AlphaFold 2: A 50-year biology problem

The claim, narrowly stated

On 30 November 2020, the organisers of the fourteenth Critical Assessment of Structure Prediction (CASP14) reported that a system built by DeepMind had predicted the three-dimensional shapes of proteins from their amino-acid sequences with an accuracy that, for many targets, approached what laboratories measure experimentally. DeepMind's own announcement put a single figure on the result: AlphaFold reached a median score of 92.4 GDT across all targets, with a median of 87.0 GDT on the hardest "free-modelling" category, where no closely related structure existed to lean on. GDT, the Global Distance Test, runs from 0 to 100 and counts how many of a model's atoms sit within increasing distance thresholds of the true positions. John Moult, who co-founded CASP in 1994, had long treated a score near 90 as informally competitive with experiment. These are task-specific figures on a blind assessment, not a measure of biological understanding and certainly not a percentage of human intelligence.

The precise event is worth separating from the mythology that grew around it. What was demonstrated was a large, reproducible improvement on one well-defined problem: predicting the folded shape of a single protein chain from its sequence, judged against structures the entrants had never seen. That is distinct from "solving biology," and distinct again from the later peer-reviewed method paper and the public database, which were separate steps taken in 2021. Keeping those layers apart is the only honest way to say why the result mattered.

Anfinsen's dogma and the fifty-year framing

The "50-year-old grand challenge" language is not marketing hyperbole; it points to a specific scientific lineage. In his 1972 Nobel lecture, Christian Anfinsen argued from ribonuclease refolding experiments that the amino-acid sequence of a protein should contain all the information needed to specify its native structure. If sequence determines shape, then shape should in principle be computable from sequence alone. The trouble was that nobody knew the function mapping one to the other. Cyrus Levinthal had already sharpened the difficulty in 1969: a modest protein has an astronomical number of possible conformations, far too many to search by brute force, yet real proteins fold reliably in microseconds to seconds. Folding is evidently directed, but the rules governing it remained opaque.

The practical consequence was a widening gap. Experimental methods (X-ray crystallography, nuclear magnetic resonance, and later cryo-electron microscopy) are slow and expensive, and by 2020 the Protein Data Bank held fewer than 200,000 solved structures. Meanwhile sequence databases catalogued hundreds of millions of proteins whose shapes were unknown. The constraint AlphaFold 2 addressed was not a lack of theory (Anfinsen's hypothesis was decades old) but the absence of a reliable computational route from sequence to coordinates for proteins with no experimental template.

CASP: the blind test that made the claim credible

The reason the 2020 result carried weight, rather than reading as another self-reported benchmark, is the venue. CASP is a biennial community experiment founded by Moult and Krzysztof Fidelis in 1994. Its design is adversarial in the good sense: organisers release the sequences of proteins whose structures have been solved experimentally but not yet published, participants submit predictions blind, and independent assessors score the predictions once the true structures are released. No entrant can tune to the answer because the answer is withheld until after submission.

That structure matters for interpreting the event. Many machine-learning claims rest on benchmarks the authors chose and can overfit. CASP removes that degree of freedom. When DeepMind's group posted top scores across roughly 100 domains in CASP14, and did so by a margin over the field that Moult described in unusually strong language, the assessment itself was the evidence. It is why the narrow claim ("best-in-class blind structure prediction, near experimental accuracy on many targets") is well supported, even as broader claims deserve more caution.

What AlphaFold 1 had already shown at CASP13

AlphaFold 2 did not arrive from nowhere. Two years earlier, at CASP13 in 2018, an earlier DeepMind system also won its category. That version, described in a 2020 Nature paper by Senior and colleagues, worked in two stages: a deep residual convolutional network predicted a distribution over pairwise distances between residues (a "distogram"), and a separate optimisation step then folded a chain to satisfy that predicted geometry using gradient descent on a potential. It was the strongest method at CASP13, yet clearly short of experimental accuracy, and its two-stage design meant the structural step could not fully inform the learning step.

The predecessor matters because it locates the real advance. AlphaFold 1 established that deep learning on evolutionary data could beat physics-based and fragment-assembly approaches at contact and distance prediction. What changed for CASP14 was architectural: the second system was rebuilt as a single, end-to-end differentiable model that reasons about geometry internally rather than bolting a folding step onto a predicted distance map. The two-year gap between a strong result and a near-solved-task result is itself informative about how the field moved.

Inside AlphaFold 2: evolution, pairs, and geometric attention

The method was published in July 2021 in Nature, in Jumper and colleagues' paper on highly accurate protein structure prediction. Its core ideas are worth stating plainly. AlphaFold 2 begins not from a single sequence but from a multiple sequence alignment: the many evolutionary relatives of the target protein. Positions that mutate in a correlated way across evolution tend to be in physical contact, because a destabilising change at one site is often compensated by a change at its neighbour. Exploiting that coevolutionary signal was already established (see the competing approaches below); AlphaFold 2's contribution was how thoroughly it integrated it.

The network maintains two representations that talk to each other through attention: one for the sequence alignment and one for every pair of residues. This "Evoformer" stack repeatedly exchanges information between the evolutionary view and the geometric, pairwise view, and it nudges the pair representation toward geometric self-consistency (the predicted distances between three residues, for instance, should respect the triangle inequality). A separate structure module then converts these abstract representations into explicit atomic coordinates using a form of geometric attention that operates directly in three dimensions, treating each residue as a movable frame. The whole system is trained end to end against a structural loss, iterated through a small number of "recycling" passes that let it refine its own output, and bootstrapped with self-distillation on a large set of predicted structures for sequences without experimental data. Crucially, it also emits a per-residue confidence score (pLDDT, scaled 0 to 100) that turns out to be well calibrated, so users can tell which parts of a prediction to trust.

The numbers, kept task-specific

Precision here guards against overclaiming. On the CASP14 targets, the Nature paper reports a median backbone accuracy of 0.96 Angstroms r.m.s.d.95 (root-mean-square deviation over the best-aligned 95% of residues), against roughly 2.8 Angstroms for the next best method, and an all-atom accuracy near 1.5 Angstroms. For context, the width of a carbon atom is about 1.5 Angstroms, so the typical backbone error was smaller than an atom's diameter on many targets. The headline 92.4 median GDT and the 87.0 GDT on template-free targets, reported in DeepMind's CASP14 announcement, are the same result expressed in CASP's own metric.

Each of these is a statement about a specific, blindly assessed task, not a global verdict. A high median leaves a tail of harder cases; GDT and r.m.s.d. measure agreement of coordinates, not whether a predicted pocket binds a particular drug; and the confidence metric, useful as it is, is a prediction about a prediction. Reporting the figures as what they are (accurate coordinates on many single-chain targets) is more defensible, and more useful, than translating them into a claim that folding is "done."

The competing approaches of its moment

AlphaFold 2 did not compete in a vacuum. Three alternative traditions defined the field it entered. The first was physics-based simulation, attempting to fold proteins by modelling atomic forces directly, an approach with deep validity but forbidding compute cost that had struggled to reach experimental accuracy on novel folds. The second was coevolution and statistical contact prediction, which read correlated mutations in sequence alignments to infer which residues touch; methods in this family made contact prediction genuinely useful in the 2010s. The third, and closest rival, fused coevolution with deep learning, exemplified by David Baker's laboratory. Their trRosetta method, published in PNAS in 2020, predicted inter-residue orientations as well as distances and folded structures with the Rosetta framework, and it was a leading approach going into CASP14.

The Baker group's response afterward is part of why the event holds up. In July 2021 they published RoseTTAFold, a three-track neural network, in Science, which independently reached accuracy approaching AlphaFold 2's and was released as open software. That independent reproduction, from a rival lab using a related but distinct architecture, is strong corroboration that the underlying result was real and not an artefact of one team's setup. Competing approaches are the right yardstick for the 2020 event: the advance was decisive relative to trRosetta and physics-based folding, and it was quickly matched in kind by a peer.

From a contest result to a database of the proteome

The downstream influence became concrete when prediction turned into infrastructure. In July 2021, alongside the method paper, DeepMind and EMBL's European Bioinformatics Institute launched the AlphaFold Protein Structure Database, initially covering the human proteome and 20 model organisms. In 2022, described in Varadi and colleagues' paper in Nucleic Acids Research, the database expanded to over 200 million predicted structures, spanning nearly every catalogued protein in the UniProt reference database. This dissemination step, distinct from the CASP result itself, is what put predicted structures into the daily workflow of biologists worldwide, from studying disease mutations to characterising previously uncharacterised proteins in neglected pathogens.

It is important to keep the dissemination separate from the science. The database massively lowered the cost of getting a plausible starting structure, but a predicted model in a database is still a hypothesis. Its value came from being both large in coverage and honest about confidence, so that researchers could triage which predictions to trust and which to treat with care. That combination, scale plus calibrated uncertainty, is arguably as consequential for practice as the accuracy figure that made the news.

The honest limits: a structure is not a protein

A predicted static structure is not the full behaviour of a protein, and the AlphaFold authors were explicit about this. The system predicts a single, high-confidence conformation, typically the folded native state, but many proteins are functional precisely because they move: they flex, switch between states, and respond to binding partners. AlphaFold 2 does not by itself predict those conformational ensembles, folding pathways, or kinetics. It was trained on single chains, so predicting how proteins assemble into complexes required a separate adaptation (AlphaFold-Multimer), and it does not natively model small-molecule ligands, ions, or nucleic acids. Regions that are intrinsically disordered are correctly flagged with low confidence rather than resolved. And because the model learned from the Protein Data Bank, genuinely novel folds unlike anything in that record remain a real source of error.

There is also a subtler caveat about the word "solved." What advanced dramatically was single-domain, single-chain structure prediction under blind assessment. Predicting how a mutation changes stability or function, how a drug binds, or how a structure behaves in a cell are related but different problems, several still open. A high GDT does not certify a validated drug target or a mechanistic explanation. Treating the result as a powerful new instrument, one that still requires experimental confirmation for consequential claims, is the accurate framing.

Recognition, and the successors it enabled

The scientific community's judgment arrived formally in 2024, when the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry with one half to Demis Hassabis and John Jumper for protein structure prediction and the other half to David Baker for computational protein design. A Nobel awarded within four years of the CASP14 result is unusually fast, and the shared citation is itself instructive: it recognises both prediction (reading structure from sequence) and design (writing new sequences for desired structures) as two faces of the same computational advance.

The line of work also continued. In 2024, DeepMind and Isomorphic Labs published AlphaFold 3 in Nature, described in an accompanying DeepMind announcement, extending the approach from single proteins to complexes involving nucleic acids, ligands, and ions using a diffusion-based architecture. That successor addresses several of the 2020 system's limits directly, which is exactly what a healthy research trajectory looks like: a landmark result followed by work that names its boundaries and pushes past them.

How to place the event in AI history

The defensible reading is that CASP14 was a landmark demonstration, not a total resolution. Its narrow, well-evidenced claim is that a deep-learning system produced near-experimental blind predictions of single-chain protein structures, a task with a fifty-year pedigree and a rigorous, adversarial assessment. Its broader significance comes from placing that result in context: it showed that machine learning could deliver a genuinely scientific advance outside its origins in language and vision, and that domain-specific structure (evolutionary alignments, geometric reasoning, calibrated confidence) mattered as much as raw scale.

A useful way to navigate outward from this event is to hold three questions apart. What earlier idea supplied a necessary ingredient (Anfinsen's hypothesis, coevolution, the distogram approach of AlphaFold 1)? What independent work corroborated the result (RoseTTAFold from a rival lab)? What limitation demanded further discovery (complexes, ligands, dynamics, addressed later and still not fully)? Answered that way, the story avoids both extremes: it neither shrinks a real breakthrough into an incremental step nor inflates it into a claim that biology was finished on 30 November 2020. The event reorganised what structural biologists could routinely attempt, and that is a large enough thing to say without embellishment.

References

• Google DeepMind (2020). AlphaFold: a solution to a 50-year-old grand challenge in biology.
• Jumper et al. (2021). Highly accurate protein structure prediction with AlphaFold.
• CASP organizers (2020). CASP14: 14th Community Wide Experiment on the Critical Assessment of Structure Prediction.
• Senior et al. (2020). Improved protein structure prediction using potentials from deep learning.
• Baek et al. (2021). Accurate prediction of protein structures and interactions using a three-track neural network.
• Yang et al. (2020). Improved protein structure prediction using predicted interresidue orientations.
• Varadi et al. (2022). AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space.
• EMBL-EBI and Google DeepMind (2021). AlphaFold Protein Structure Database.
• Anfinsen (1972). Studies on the principles that govern the folding of protein chains.
• The Royal Swedish Academy of Sciences (2024). The Nobel Prize in Chemistry 2024.
• Abramson et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3.
• Google DeepMind (2024). AlphaFold 3 predicts the structure and interactions of all of life's molecules.