The AlphaFold Protein Folding Breakthrough

In November 2020, a computer program did something biologists had chased for half a century: it predicted the three-dimensional shape of a protein from nothing but the sequence of amino acids that build it, and it did so accurately enough that its answers were nearly indistinguishable from painstaking lab measurements. The program, AlphaFold, built by Google DeepMind, entered a blind global competition called CASP14 and beat every other method by a wide margin. Four years later, in October 2024, its creators John Jumper and Demis Hassabis shared the Nobel Prize in Chemistry with David Baker for the work.

The 50-year problem AlphaFold solved

A protein is a chain of amino acids, but the chain by itself does almost nothing. To function, that chain has to fold into a precise 3D shape, and the shape determines everything: whether an enzyme can grab its target, whether an antibody can bind a virus, whether a drug molecule fits into a receptor. Misfolded proteins are directly implicated in diseases including Alzheimer's, Parkinson's, and cystic fibrosis.

The trouble is that figuring out the shape experimentally, through techniques like X-ray crystallography or cryo-electron microscopy, is slow and expensive, often taking months or years and costing tens of thousands of dollars per protein. By contrast, sequencing the amino acid chain itself is cheap and fast. That gap produced a massive backlog: by 2020, scientists had sequenced roughly 200 million known proteins but experimentally solved the structure of only a small fraction of them.

The theoretical version of this problem has a name, Levinthal's paradox, after molecular biologist Cyrus Levinthal, who pointed out in 1969 that a typical protein chain has an almost incomprehensible number of possible 3D configurations, on the order of 10^300 by his estimate (Wikipedia, summarizing Levinthal 1969). Trying every possible fold by brute-force calculation would take far longer than the age of the universe, yet real proteins in a living cell fold into their correct shape in milliseconds. Something about the physics and chemistry of the amino acid chain guides it directly to the right answer, and for decades no computational method could reliably reverse-engineer what that something was.

Levinthal estimated that a typical protein has on the order of 10^300 possible folded configurations. Enumerating them one by one, even at an impossibly fast rate, would take vastly longer than the roughly 13.8-billion-year age of the universe. Real proteins fold correctly in milliseconds ([Wikipedia](https://en.wikipedia.org/wiki/Levinthal's_paradox)).

How AlphaFold actually predicts a shape

AlphaFold does not simulate the physical folding process the way a molecular dynamics program would, letting atoms jostle step by step toward a stable shape. Instead, it treats structure prediction as a pattern-recognition problem, trained on the roughly 170,000-plus experimentally solved structures sitting in the Protein Data Bank, the public repository researchers have built up since the 1970s.

Given a new amino acid sequence, AlphaFold first searches large sequence databases for related proteins across many species, building what's called a multiple sequence alignment (a side-by-side comparison of similar sequences from evolutionarily related proteins). Because evolution tends to preserve amino acid pairs that sit physically close together in the folded structure, even when they're far apart in the raw sequence, patterns of co-evolution in that alignment carry hidden clues about which parts of the chain touch which other parts once folded.

AlphaFold's core neural network, called Evoformer, processes that evolutionary and structural information together and iteratively refines a 3D structure prediction, passing information back and forth between the sequence-relationship representation and the spatial-geometry representation until both converge on a consistent shape. The output is not just a single guess but a full atomic coordinate model, plus a confidence score (called pLDDT) for how reliable each part of the structure is likely to be.

How accurate is AlphaFold, exactly?

At CASP14 in 2020, AlphaFold2 achieved a median score of about 92.4 on the competition's Global Distance Test (a 0–100 accuracy scale used to grade predicted protein structures against the real, experimentally solved ones). That score corresponds to a median error of under 1 angstrom, roughly the width of a single atom, putting AlphaFold's predictions in the same accuracy range as experimental methods for many targets and about three times more accurate than the next-best competing method (UC Davis Genome Center; Nature).

Why the result mattered beyond one competition

CASP, the Critical Assessment of Structure Prediction, exists specifically to prevent researchers from grading their own homework. Every two years, participating teams receive amino acid sequences for proteins whose structures have already been solved experimentally but not yet published, predict the shape blind, and submit before the real answer is revealed. That blind, adversarial format is why AlphaFold's CASP14 result carried real scientific weight rather than being a claim DeepMind made about its own model.

DeepMind's decision to open-source AlphaFold2's code and, working with EMBL's European Bioinformatics Institute, publish a searchable database of predicted structures turned a single lab's tool into shared global infrastructure. The AlphaFold Protein Structure Database has grown from an initial release of under 1 million structures in 2021 to more than 214 million predicted protein structures by 2024, effectively covering nearly every protein sequence known to science (Nucleic Acids Research).

That scale changed what was practical for individual research groups. A lab studying a rare disease-linked protein that would previously have needed years and a dedicated crystallography setup to get a structure can now, in many cases, look the predicted structure up directly or generate one in hours. Researchers have used AlphaFold-derived structures to accelerate work on drug target identification, enzyme design, and understanding disease-linked mutations, and the tool has become a standard reference alongside experimental methods rather than a replacement fringe curiosity.

The 2024 Nobel Prize in Chemistry was split three ways: half to David Baker "for computational protein design," and the other half jointly to Demis Hassabis and John Jumper "for protein structure prediction," recognizing AlphaFold specifically ([Nature](https://www.nature.com/articles/d41586-024-03214-7)).

What AlphaFold does not do

AlphaFold predicts a static, single "best guess" structure extremely well for most proteins, but plenty of important biology happens outside that frame. Many proteins are intrinsically disordered, meaning they don't settle into one fixed shape at all but instead shift between conformations depending on their environment, and AlphaFold's confidence scores for those regions are typically, and correctly, low rather than falsely precise.

The model is also weaker on predicting how a protein's shape changes when it interacts with drugs, other molecules, or changing cellular conditions, and on modeling large, dynamic protein complexes with many moving parts. Later versions, including AlphaFold3 released in 2024, extended prediction to include how proteins interact with DNA, RNA, and small molecules, but this remains an active and less mature area than single-protein structure prediction (PMC).

What AlphaFold does well	What remains a limitation
Predicting single-chain protein structure	Modeling intrinsically disordered regions
Matching experimental accuracy for many targets	Capturing dynamic conformational changes
Scaling to hundreds of millions of proteins	Predicting drug-binding interactions reliably
Providing per-residue confidence scores	Modeling large multi-protein complexes fully

Table: AlphaFold's core strengths versus its documented limits, based on CASP results and peer-reviewed follow-up analyses cited throughout this article.

Researchers examining AlphaFold critically have also pointed out that it has not solved Levinthal's paradox in the sense of explaining the underlying physics of how folding happens step by step. It predicts the end state of folding with remarkable accuracy without describing the physical folding pathway a protein takes to get there. Structural biologists distinguish this clearly: AlphaFold answered "what shape will this protein take," not "how does the folding process itself unfold in real time" (FASEB Journal).

What came before AlphaFold, and why it fell short

AlphaFold2's 2020 result did not appear in a vacuum. Structural biologists had spent decades building computational workarounds for the fact that experimental structure determination was slow. The main pre-deep-learning approaches were homology modeling, which builds a new structure by copying the known shape of a related, already-solved protein and adjusting for sequence differences, and ab initio (or "from scratch") modeling, which tries to predict a fold using physics-based energy calculations without relying on a known template at all.

Tools like Swiss-Model automated homology modeling into a streamlined pipeline, and Rosetta, developed by David Baker's lab, became the leading ab initio method, competing at CASP as far back as its fourth edition in 2000. Homology modeling worked reasonably well when a close relative's structure was already known, but it failed for proteins with no good template, which described a large share of the unsolved structural universe. Ab initio methods like Rosetta could tackle unfamiliar folds but were computationally expensive and far less accurate, especially on larger or more complex proteins (PMC, "Before and after AlphaFold2").

DeepMind's first version, AlphaFold1, entered CASP13 in 2018 and already outperformed the field, but it still leaned heavily on template-based and homology-style techniques. AlphaFold2 was a structurally different system that no longer depended primarily on finding a related, already-solved template. At CASP14, this showed up as a stark numerical gap: AlphaFold2 posted a median backbone accuracy (measured by RMSD, root-mean-square deviation, a standard error measurement in angstroms) of about 0.8 angstroms, compared to roughly 2.8 angstroms for the next-best method, and a summed z-score of 244.0 against 90.8 for the second-place group, a margin so large that CASP's own organizers described it as a result they had not expected to see for years, if ever (ScienceDirect, "Protein Structure Prediction: Challenges, Advances, and the Shift of Research Paradigms"). It's worth noting the top server-based group at CASP14 was BAKER-ROSETTASERVER, running a deep-learning-enhanced descendant of Rosetta called trRosetta, meaning even the best available alternative to AlphaFold2 was already borrowing the co-evolutionary, deep-learning ideas AlphaFold had helped popularize, and it still trailed by a wide margin.

AlphaFold 3: from one protein to entire molecular systems

AlphaFold2 answered a specific question well: given one protein's sequence, what shape does it fold into. Real biology, however, mostly happens through interactions, a protein binding a strand of DNA, an enzyme grabbing a small-molecule drug, two proteins locking together to form a working complex. AlphaFold3, released by Google DeepMind and its drug-discovery spinoff Isomorphic Labs in May 2024, was built specifically to model those interactions rather than single isolated chains.

The architectural shift underneath AlphaFold3 is substantial. AlphaFold2's core engine, Evoformer, relied heavily on a deep stack of 48 processing blocks built around the multiple sequence alignment, the side-by-side comparison of related sequences across species described earlier in this article. AlphaFold3 replaces Evoformer with a leaner network called Pairformer that cuts the MSA-processing stack down to just four blocks, shifting more of the modeling weight onto direct pairwise relationships between structural elements. More importantly, AlphaFold3 swaps out AlphaFold2's deterministic structure module, which computed one direct, non-generative answer, for a diffusion-based generative model, the same family of technique that powers image generators like DALL-E. Instead of computing coordinates in one pass, the diffusion module starts from randomized, noisy atomic positions and iteratively denoises them across many steps until they settle into a coherent 3D structure (Medium, "AlphaFold3 and its improvements in comparison to AlphaFold2"; Exxact).

That architecture change is what unlocked AlphaFold3's biggest practical upgrade: it is multi-modal, meaning it was trained to jointly predict structures involving not just proteins but also DNA, RNA, small molecules, ions, and chemically modified residues, all interacting with each other in a single predicted complex. AlphaFold2 could only ever hand back a protein's shape in isolation. AlphaFold3 can, for instance, model how a candidate drug molecule physically sits inside a protein's binding pocket, which is far closer to what a pharmaceutical chemist actually needs to evaluate a compound. Because generative diffusion models can occasionally produce a plausible-looking but physically wrong structure, a failure mode researchers call hallucination, DeepMind trained AlphaFold3 with a cross-distillation process against AlphaFold-Multimer and AlphaFold2 outputs specifically to reduce that risk, though independent reviewers still flag structural hallucination and complex-modeling reliability as open concerns worth watching (Deep Origin, "AlphaFold 3: Exciting Advance yet Unresolved Major Issues Remain").

What's the practical difference between AlphaFold2 and AlphaFold3 for a working scientist?

AlphaFold2 tells you the shape of one protein chain. AlphaFold3 can model that same protein physically interacting with a drug candidate, a strand of RNA, or another protein, in one combined prediction. That shift moves AlphaFold from a structural biology reference tool toward something closer to a drug-design engine, which is why Isomorphic Labs, not just academic labs, is one of its primary users.

From prediction to the pharmacy: how AlphaFold is used in real drug discovery

The gap between "we can predict a protein's shape" and "we can develop a new medicine" is enormous, and AlphaFold does not close it by itself. What it does is remove one of the slowest, most expensive early steps, getting a reliable structure for the target protein a drug needs to act on, and that alone has changed the economics of early-stage drug research for academic groups and biotech companies that could never afford months of crystallography for every candidate target.

Isomorphic Labs, the Alphabet-owned drug discovery company spun out of DeepMind, was built specifically to take AlphaFold's structural predictions and extend them into a full computational drug-design pipeline, and it has since formed research partnerships with major pharmaceutical companies to apply AlphaFold3-derived modeling to real drug candidates rather than academic case studies (Isomorphic Labs, "Our Tech"; Google Blog, "Google DeepMind and Isomorphic Labs introduce AlphaFold 3 AI model").

Beyond that flagship spinout, the freely available AlphaFold Protein Structure Database has been picked up broadly across neglected-disease research, an area historically underfunded because market-driven pharmaceutical R&D has little financial incentive to develop treatments for diseases concentrated in low-income regions. Google DeepMind has specifically cited use of AlphaFold2 by partners researching leishmaniasis and Chagas disease, along with work on malaria vaccine targets, antibiotic resistance, and structural work related to SARS-CoV-2, the virus behind COVID-19 (Google DeepMind, "AlphaFold"). Separately, academic researchers used AlphaFold to computationally screen candidate drug targets in Trypanosoma cruzi, the parasite that causes Chagas disease, illustrating exactly the kind of neglected-pathogen research that previously moved slowly due to the cost of experimentally solving each target's structure (PMC, "The Use of AlphaFold for In Silico Exploration of Drug Targets in Trypanosoma cruzi").

It's worth being precise about what this means in practice. AlphaFold accelerates target identification and candidate screening, the earliest, cheapest-to-fail stages of drug development. It does not run clinical trials, does not guarantee a predicted binding interaction will translate into a safe, effective drug in a living patient, and does not replace the years of toxicology, dosing, and trial work that still stand between a promising computational lead and an approved medicine. Every drug discovery program that uses AlphaFold-derived structures still has to clear the same regulatory and clinical bar as one that started from an X-ray crystal structure.

Common misconceptions about AlphaFold

The most common misreading treats AlphaFold's predictions as experimentally verified facts rather than model outputs with associated confidence scores. A high-confidence region of a predicted structure is genuinely reliable and regularly matches experimental structures closely, but AlphaFold itself flags lower-confidence regions, and treating every predicted structure as ground truth ignores information the tool is explicitly providing.

A second misconception is that AlphaFold has made experimental structural biology, X-ray crystallography and cryo-electron microscopy, obsolete. In practice, experimental methods remain essential both for validating predictions and for the categories of structures AlphaFold handles poorly: large dynamic complexes, unusual folds absent from its training data, and structures involving drug binding or unusual cellular conditions. AlphaFold accelerates and extends structural biology; it does not replace the lab bench.

A third misconception, common in general coverage of the story, is that AlphaFold "solved" protein folding as a physics problem. What it solved is the prediction problem: given a sequence, output the likely shape. Levinthal's original question, how the physical folding process itself manages to avoid searching an astronomical number of configurations, is still an open area of biophysics research.

The takeaway

AlphaFold's breakthrough was not a single dramatic experiment but a validated, blind-tested leap in prediction accuracy that turned a decades-long computational bottleneck into a largely solved engineering problem for most proteins. Its practical impact came less from the CASP14 scoreboard and more from what DeepMind did afterward: opening the model and publishing a freely searchable database now covering over 214 million predicted structures. The honest caveat that belongs in every account of the story is that AlphaFold predicts shape with extraordinary accuracy without explaining the physical mechanism of folding itself, and it still struggles with the flexible, dynamic, and interacting proteins that make up a large share of real biology.

---

Read next:

• The Double-Slit Experiment: How Light Proved the Universe Is Weird
• Quantum Mechanics Foundations Explained
• How the Brain Works: Neurons, Regions, and the Mind

Frequently Asked Questions

Q: What is AlphaFold? A: AlphaFold is an AI system built by Google DeepMind that predicts a protein's 3D structure directly from its amino acid sequence, using a neural network trained on experimentally solved structures from the Protein Data Bank.

Q: How accurate is AlphaFold compared to lab methods? A: At the CASP14 competition in 2020, AlphaFold2 achieved a median accuracy score of about 92.4 out of 100 on the Global Distance Test, with a median error under 1 angstrom, putting it in the same range as many experimental structure-determination methods (UC Davis Genome Center).

Q: Did AlphaFold win a Nobel Prize? A: Yes. The 2024 Nobel Prize in Chemistry went jointly to Demis Hassabis and John Jumper for AlphaFold, and to David Baker for separate work on computational protein design (Nature).

Q: Does AlphaFold solve Levinthal's paradox? A: Not in the strict sense. Levinthal's paradox is about the physical mechanism proteins use to fold quickly despite an astronomical number of possible shapes. AlphaFold predicts the final folded shape with high accuracy but does not model or explain the physical folding pathway itself (FASEB Journal).

Q: What can't AlphaFold do well? A: It struggles with intrinsically disordered proteins that don't hold one fixed shape, with modeling how proteins change conformation when they bind drugs or other molecules, and with large, dynamic multi-protein complexes. Later versions like AlphaFold3 extend into some of these areas but remain less mature than single-protein prediction (PMC).

Q: What is the difference between AlphaFold2 and AlphaFold3? A: AlphaFold2 predicts the shape of a single protein chain in isolation. AlphaFold3, released in 2024, uses a diffusion-based generative architecture and can jointly model proteins interacting with DNA, RNA, small molecules, and drug candidates, making it far more useful for drug design work (Medium, "AlphaFold3 and its improvements in comparison to AlphaFold2").

Q: What methods did scientists use to predict protein structures before AlphaFold? A: The two dominant pre-deep-learning approaches were homology modeling, which copies the known shape of a related, already-solved protein, and ab initio modeling, which uses physics-based energy calculations to build a structure from scratch. Tools like Swiss-Model and Rosetta led these approaches, but both were far less accurate and much slower than AlphaFold2 (PMC, "Before and after AlphaFold2").

Q: Has AlphaFold actually been used to discover real drugs? A: AlphaFold's structural predictions have been used to accelerate early-stage drug target identification, including work on neglected diseases like leishmaniasis and Chagas disease, and malaria vaccine research. Isomorphic Labs, a DeepMind spinout, uses AlphaFold3-derived modeling in partnership with pharmaceutical companies, but AlphaFold itself does not run clinical trials or replace the years of safety and efficacy testing required before a drug is approved (Isomorphic Labs; Google DeepMind).

Q: How much more accurate was AlphaFold2 than the next-best method at CASP14? A: AlphaFold2 achieved a median backbone accuracy of about 0.8 angstroms of error, compared to roughly 2.8 angstroms for the next-best method, and a summed z-score of 244.0 versus 90.8 for the second-place group, a margin CASP organizers described as far beyond what they expected to see (ScienceDirect).

Sources

• Chemistry Nobel goes to developers of AlphaFold AI that predicts protein structures, Nature
• UC Davis Genome Center Spotlight: CASP and the Nobel Prize-Winning Breakthrough of AlphaFold
• AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences, Nucleic Acids Research
• Levinthal's paradox, Wikipedia summary of Levinthal 1969
• Protein structure: Has Levinthal's paradox "folded"?, The FASEB Journal
• The transformative impact of AI-enabled AlphaFold 3, PMC
• AlphaFold — Google DeepMind
• Before and after AlphaFold2: An overview of protein structure prediction, PMC
• Protein Structure Prediction: Challenges, Advances, and the Shift of Research Paradigms, ScienceDirect
• AlphaFold3 and its improvements in comparison to AlphaFold2, Medium
• AlphaFold 3 Update, Exxact Blog
• AlphaFold 3: Exciting Advance yet Unresolved Major Issues Remain, Deep Origin
• Google DeepMind and Isomorphic Labs introduce AlphaFold 3 AI model, Google Blog
• Isomorphic Labs — Our Tech
• The Use of AlphaFold for In Silico Exploration of Drug Targets in Trypanosoma cruzi, PMC