AlphaFold: How AI Solved a 50-Year Protein Folding Problem

In November 2020, an AI system called AlphaFold made a scientific announcement so significant that the Nobel Committee would eventually recognise it with the highest prize in chemistry. DeepMind’s AlphaFold had cracked one of biology’s grand challenges: the protein folding problem. It had solved, with extraordinary accuracy, a puzzle that had occupied scientists for fifty years and that many thought would not be solved in their lifetimes. This article explains what protein folding is, why it matters enormously for medicine, how AlphaFold works, and what it has already begun to change.

What Is a Protein and Why Does Its Shape Matter?

To understand protein folding, you first need to understand proteins. Every living cell — in your body, in a bacterium, in a plant — is run by proteins. They are the molecular machines that do virtually all the work of life. Enzymes are proteins that catalyse chemical reactions, speeding them up by factors of millions or billions. Antibodies are proteins that recognise and neutralise pathogens. Receptors are proteins that sit on cell surfaces and detect signals from the outside world. Haemoglobin is a protein that carries oxygen in your blood. Collagen is a protein that gives structural integrity to your skin, tendons, and bones.

Proteins are made of chains of smaller molecules called amino acids. There are twenty different amino acids, and a typical protein might contain anywhere from a few dozen to several thousand of them, strung together in a specific sequence. That sequence is encoded in your DNA — your genome is, in a very real sense, a recipe book for proteins.

But a protein is not just a long string. As soon as the chain is synthesised, it folds up into a precise, specific three-dimensional shape — and that shape is what determines what the protein does. A haemoglobin molecule that has folded incorrectly cannot carry oxygen. An enzyme with the wrong shape cannot catalyse its target reaction. Many diseases — including Alzheimer’s, Parkinson’s, cystic fibrosis, and many cancers — are caused or worsened by proteins that fold into the wrong shapes. Understanding a protein’s function requires knowing its shape. And figuring out what shape a protein will fold into, given only its amino acid sequence, is the protein folding problem.

For a beginner-friendly overview of AI breakthroughs like this one, see our guide to artificial intelligence.

Why Was Protein Folding So Hard?

In principle, protein folding is a physics problem. The laws of chemistry dictate how atoms interact, and in principle you could calculate the lowest-energy conformation — the folded shape — of any protein by simulating the motion of every atom. In practice, this has been computationally impossible for any protein of realistic size. A medium-sized protein with 300 amino acids has hundreds of thousands of atoms, and the number of possible conformations it could adopt is astronomically large.

In 1969, Cyrus Levinthal pointed out a famous paradox: if a protein tried every possible conformation at random to find its lowest-energy state, it would take longer than the age of the universe. Yet real proteins fold into their correct shape in microseconds to seconds. Something else must be going on — evolution has shaped the amino acid sequence so that the folding process follows a directed pathway rather than a random search — but figuring out what that pathway is for any given sequence remained beyond scientists’ capabilities.

Experimental methods exist for determining protein structures: X-ray crystallography, cryo-electron microscopy (cryo-EM), and nuclear magnetic resonance (NMR) spectroscopy can all resolve protein structures at atomic resolution. But these methods are slow, expensive, and technically demanding. Growing a crystal of a protein for X-ray crystallography can take months or years and fails entirely for many proteins, particularly membrane proteins. As of 2020, the Protein Data Bank — the global repository for known protein structures — contained structures for roughly 170,000 proteins. But life uses hundreds of millions of different proteins. The gap between what was known and what was needed was immense.

The CASP Competition: Measuring Progress

Since 1994, the biennial Critical Assessment of Structure Prediction (CASP) competition has provided the benchmark for protein structure prediction. Organisers select a set of proteins whose structures have been determined experimentally but not yet publicly released. Competing teams submit their predicted structures, which are then compared against the experimental results. Performance is measured using a metric called the Global Distance Test (GDT) score, on a scale of 0 to 100, where 100 means a perfect match to the experimentally determined structure.

A GDT score of around 90 is considered comparable to experimental accuracy. For decades, the best methods struggled to reach scores in the 40s and 50s on difficult targets. Progress was real but slow, measured in increments over years. Then came CASP14 in 2020.

DeepMind entered CASP14 with AlphaFold 2, a complete redesign of its earlier AlphaFold system. The results were so far beyond anything that had been seen before that they left the structural biology community stunned. AlphaFold 2 achieved a median GDT score of 92.4 across all targets, and on many individual proteins it matched experimental structures to within the margin of error of the experimental methods themselves. It had solved the protein folding problem — at least to a first approximation. John Moult, a founder of CASP, described it as ‘a stunning advance.’ AlQuraishi, a prominent researcher in the field, called it ‘the most significant achievement in structural biology since the determination of the DNA double helix.’

For more context on AI milestones like CASP14, see our complete history of artificial intelligence.

How AlphaFold 2 Works

AlphaFold 2, published in Nature in July 2021 alongside its source code and the structures of the human proteome (every protein encoded in the human genome), represents a genuinely novel approach to the problem. It is built on the Transformer architecture that has also powered large language models, adapted for biological sequences.

The system takes as input the amino acid sequence of a protein and, crucially, a multiple sequence alignment (MSA) — a comparison of that protein’s sequence with related sequences from hundreds or thousands of other species. Evolution acts as a signal: if two positions in a protein sequence have co-evolved across millions of years (changing together in different species), they are likely to be physically close in three-dimensional space, because mutations that would disrupt an important contact are selected against. AlphaFold learns to read these evolutionary signals with extraordinary sensitivity.

The network processes the sequence and the evolutionary data through a series of Transformer-like ‘Evoformer’ blocks, which iteratively refine a representation of the relationships between all pairs of amino acid residues. This representation is then fed into a structure module that predicts the three-dimensional coordinates of every atom in the protein. The entire process, which might take an experimental structural biologist years, runs in minutes on a modern computer.

DeepMind trained AlphaFold 2 on the entire Protein Data Bank, using roughly 170,000 known structures alongside larger databases of protein sequences. The training required substantial computing resources, but the resulting system can run predictions on consumer hardware.

More details on DeepMind’s work can be found in our article on AlphaGo, the earlier breakthrough from the same company.

The Nobel Prize and the AlphaFold Database

In 2024, Demis Hassabis, the co-founder and CEO of DeepMind, was awarded the Nobel Prize in Chemistry alongside John Jumper, the lead researcher on AlphaFold 2, and David Baker, a structural biologist who independently pioneered computational protein design. The Nobel Committee cited AlphaFold’s role in solving ‘a fifty-year-old grand challenge of biology.’ It was an extraordinary validation of AI as a tool for fundamental scientific discovery — and the first Nobel Prize awarded primarily for the development of an AI system.

Even before the Nobel Prize, the scientific community had begun to transform how it worked. DeepMind and the European Bioinformatics Institute (EMBL-EBI) launched the AlphaFold Protein Structure Database in July 2021 with structures for the entire human proteome — over 20,000 proteins. By 2022, they had expanded the database to cover over 200 million proteins from virtually every organism that had been sequenced — essentially the entire known universe of proteins. The database is freely accessible to any researcher anywhere in the world.

The impact was immediate. Research that would have taken years of experimental structural determination could now begin with a high-confidence predicted structure from AlphaFold in a matter of minutes. Papers began appearing that used AlphaFold structures as starting points for everything from fundamental mechanistic studies of how proteins work to the design of new drugs.

AlphaFold 3 and Expanding Capabilities

In 2024, DeepMind published AlphaFold 3, a significant extension of the original system. Where AlphaFold 2 focused exclusively on proteins, AlphaFold 3 can predict the structures of complexes involving proteins together with DNA, RNA, and small molecules (including potential drug compounds). This is crucial for drug discovery, because drugs typically work by binding to specific sites on proteins, and designing a drug requires understanding how it will interact with its target in three-dimensional space.

AlphaFold 3 uses a diffusion-based approach — similar to the technology behind image-generation models like Stable Diffusion — rather than the explicit coordinate prediction of AlphaFold 2, which allows it to handle the greater diversity of molecular types. It represents another step change in capability, and its release has further accelerated the pace of computational drug discovery.

You can explore the broader AI landscape in our AI glossary.

Impact on Drug Discovery and Medicine

The most profound long-term impact of AlphaFold is likely to be in drug discovery. Developing a new drug from scratch — from identifying a target protein to running clinical trials to regulatory approval — takes on average 12 to 15 years and costs upwards of two billion dollars. The failure rate is staggering: only about 1 in 10 drugs that enter clinical trials ultimately reaches patients.

One of the most time-consuming early steps in drug discovery is determining the structure of the target protein, so that chemists can design molecules that will fit into its active site and modulate its activity. AlphaFold does not eliminate the need for experimental validation, but it dramatically accelerates the early stages, allowing researchers to move from sequence to structure in minutes and to identify potential binding sites and design initial candidate molecules far more quickly than was previously possible.

Real-world applications are already emerging. Researchers have used AlphaFold structures to investigate neglected tropical diseases, for which there is insufficient commercial incentive to fund traditional expensive structural biology work. Drug companies including AstraZeneca, Pfizer, and Novartis have integrated AlphaFold into their drug discovery pipelines. Startups built specifically around AI-driven drug discovery — using AlphaFold and related tools — have attracted billions of dollars in venture capital investment.

AlphaFold has also had immediate impact in fundamental biology. Structures for thousands of proteins that had resisted all experimental determination are now available for study. Entire families of proteins — including many involved in disease — that were previously ‘dark’ (known from genome sequences but structurally uncharacterised) are now illuminated. The era of structural biology where only the best-funded labs with access to expensive equipment could determine structures is ending; the knowledge is now democratised.

10 Ways AlphaFold Is Reshaping Science Right Now

The conceptual breakthrough is well-told. The 10 ways below describe what AlphaFold and its descendants are actually changing in 2026.

1. PhD timelines compress in structural biology

What used to be a 4-year structural-biology PhD around solving one protein structure now becomes a research question about a family of proteins. The unit of inquiry expands.

2. Drug-target validation moves earlier in the pipeline

Druggability assessment used to require synthesized molecules. AlphaFold-predicted structures plus docking simulations move this earlier; failed-target pruning happens before lab work begins.

3. Antibody-engineering acceleration

Antibody design used to be experimental and laborious. AlphaFold 3 plus successor tools accelerate antibody-target-affinity prediction; iteration cycles drop from months to weeks.

4. Enzyme engineering for biomanufacturing

Engineered enzymes for plastic degradation, biofuel production, and specialty-chemical synthesis are easier to design when you can predict structure-function relationships in silico first.

5. Genetic-disease mechanism understanding

Many disease-causing mutations produce malformed proteins. AlphaFold helps visualize the structural consequence of a mutation, which informs treatment strategy.

6. Protein-protein interaction mapping at scale

Cellular biology depends on knowing which proteins bind which. AlphaFold-Multimer plus successor tools map protein-protein interactions across whole proteomes, opening cell-biology questions previously intractable.

7. Vaccine antigen design

Designing vaccine antigens that elicit the right immune response involves structural considerations. AlphaFold accelerates vaccine-development workflows from concept to candidate.

8. CRISPR design refinement

CRISPR-Cas system engineering benefits from structural prediction of guide-RNA-target-DNA interactions. Specificity improvements drop off-target effects in therapeutic applications.

9. Public-database scale matters more than ever

The AlphaFold Protein Structure Database has predicted structures for 200+ million proteins. Researchers access predictions for free; the public commons of structural data unlocks downstream research globally.

10. The next breakthrough probably comes from protein-design AI

AlphaFold predicts structure from sequence. Reverse models (RFdiffusion, ESM3) design sequences from desired structures. The de novo protein-design wave is now arriving on the foundation AlphaFold laid.

Before AlphaFold: How Researchers Coped

The protein folding problem did not prevent structural biologists from making progress before AlphaFold — it just made progress slow, expensive, and highly selective. X-ray crystallography, the dominant technique for high-resolution protein structure determination for most of the twentieth century, requires the researcher to first grow a crystal of the protein of interest. Growing a crystal is an art as much as a science: the conditions of temperature, pH, salt concentration, and precipitant must be found by trial and error, the process can take months to years, and many proteins — particularly membrane proteins, which are embedded in the cell’s lipid bilayer and are crucially important drug targets — resist crystallisation entirely.

Cryo-electron microscopy (cryo-EM), which became a mature technique in the 2010s, partially circumvented the crystallisation bottleneck by allowing structure determination on proteins in a vitrified (frozen) state without the need for crystals. The 2017 Nobel Prize in Chemistry was awarded to Jacques Dubochet, Joachim Frank, and Richard Henderson for developing cryo-EM. But even cryo-EM requires specialised, expensive instruments, considerable operator skill, and substantial computational processing time. The technical barriers meant that structure determination remained confined to well-funded labs in wealthy institutions.

Nuclear magnetic resonance (NMR) spectroscopy provides another route to protein structures, and is particularly useful for smaller proteins and for studying their dynamic behaviour. But NMR is limited to proteins below a certain size, requires expensive instruments and expertise, and produces structures only after lengthy data collection and analysis. Taken together, all available experimental methods were producing perhaps ten thousand new protein structures per year by the late 2010s — a rate that would have taken centuries to cover the full range of proteins in nature, let alone to keep up with the genomic sequencing revolution that was identifying new proteins at an ever-accelerating pace.

The Scientific Community’s Response to AlphaFold

The reception of AlphaFold within the structural biology community was initially mixed. Some researchers celebrated the breakthrough immediately and began using AlphaFold predictions in their work within weeks of the CASP14 results being announced. Others were more cautious, pointing out that predicted structures are not the same as experimentally determined ones and that AlphaFold’s predictions could embed errors that would propagate through downstream research if accepted uncritically.

These concerns were legitimate and have been partly borne out by experience. AlphaFold predictions are highly accurate for proteins where many related sequences are available from other species — the evolutionary co-variation signal is strong — but less reliable for proteins that are evolutionarily unusual or that adopt multiple different conformations in different cellular conditions. AlphaFold gives a single predicted structure, but many proteins are inherently flexible and adopt different shapes depending on what they are bound to or what the cellular environment is. This limitation is important for drug discovery, where understanding the conformational landscape of a target protein can be crucial.

Despite these caveats, the consensus view in structural biology by 2022 was that AlphaFold had genuinely transformed the field. A survey of structural biologists found that the vast majority were using AlphaFold predictions regularly, and that the predicted structures had already contributed to real discoveries. High-impact papers in journals such as Nature, Science, and Cell began routinely using AlphaFold structures as starting points or validation tools. The Protein Data Bank — which had taken decades to accumulate 170,000 experimental structures — now links every entry to the corresponding AlphaFold prediction.

The most telling measure of AlphaFold’s impact may be the sheer scale of adoption. The AlphaFold Protein Structure Database recorded over a billion queries in its first two years of operation, from researchers in over 180 countries. Proteins that had been ‘dark’ — known to exist from genomic data but structurally uncharacterised — for decades suddenly had reliable structural models available. Entire families of previously mysterious proteins — including many involved in infectious diseases and neglected tropical diseases where experimental structural biology had not been economically attractive — were illuminated overnight.

---

Frequently Asked Questions

---

Related Reading

• The Complete History of Artificial Intelligence
• AlphaGo: The AI That Beat the World Champion at Go
• What Is Artificial Intelligence? A Beginner’s Guide
• The AI Glossary: Every Term Explained Simply

---

Sources

• Grokipedia
• Stanford HAI AI Index Report

You May Also Like

• What Is Artificial Intelligence
• Best AI Tools for Beginners
• How to Use AI
• AI Tools Directory
• Best Free AI Courses