Proteins are the machines of life — enzymes, scaffolds, signals, motors, immune weapons. Every one of them works because of its shape. And for 50 years, predicting that shape from a genetic sequence alone was considered one of the hardest problems in science. Then, in 2020, a neural network solved it in an afternoon.
A protein is a chain of amino acids — small molecules strung together like beads on a string. There are 20 kinds of amino acid. That's the entire alphabet. From those 20 letters, evolution wrote every enzyme, every hormone, every muscle fiber, every antibody that has ever existed on Earth.
The linear order of amino acids, encoded directly by DNA. This alone determines everything that follows.
Nearby amino acids hydrogen-bond into regular repeating shapes: alpha-helices (spirals) and beta-sheets (pleated planks).
The entire chain folds into a compact, specific 3D structure. This shape is the protein's identity — it determines what it can bind, what it catalyzes, what it builds.
Some proteins are complexes of multiple chains working together. Hemoglobin has four. The ribosome has dozens. Each subunit folds independently, then docks.
The driving force of folding: hydrophobic collapse. Amino acids that hate water bury themselves in the centre, dragging the chain into a compact shape. Polar residues stay on the outside. This simple principle governs the folding of every protein in your body.
Proteins don't just do one thing. They are simultaneously your muscles, your enzymes, your hormones, your immune system, and the motors inside every cell. Shape determines function — and evolution has found a shape for nearly every job that needs doing.
Proteins that accelerate chemical reactions — often by factors of a million or more. Every metabolic reaction in your body is catalysed by an enzyme. Without them, life-chemistry would be too slow to support life.
The scaffolding of life. Collagen builds tendons, skin, and bone. Keratin makes hair and nails. Actin and tubulin form the internal skeleton of every cell, giving it shape and letting it move.
Proteins that bind small molecules and carry them around. Hemoglobin picks up oxygen in the lungs and drops it at tissues. Albumin ferries fatty acids through the blood. Channels in cell membranes move ions across.
Hormones are proteins (insulin, growth hormone). Receptors are proteins. The entire conversation between cells — from "grow" to "die" to "release glucose" — is proteins talking to proteins.
Antibodies are proteins shaped precisely to grab pathogens. The MHC complex is a protein that presents pathogen fragments to T-cells. Complement proteins punch holes in bacteria. The immune system is essentially a protein army.
Myosin literally walks along actin filaments, converting ATP energy into mechanical steps. This is how muscles contract. Kinesin and dynein carry cargo along microtubule highways inside cells — protein trucks on protein roads.
Every structure below is real, determined by X-ray crystallography or cryo-EM and deposited in the Protein Data Bank. Drag to rotate. Scroll to zoom. Switch proteins and view modes with the controls.
The sequence determines the shape — Christian Anfinsen proved this in 1972. But computing the shape from the sequence? That was a different matter. The search space is so vast it became its own paradox.
When a protein folds into the wrong shape, it can become toxic — aggregating into sticky clumps that damage cells. These "proteopathies" are among the hardest diseases to treat, because the culprit is structurally almost identical to the normal protein.
Amyloid-β peptides misfold and aggregate into plaques between neurons. Tau protein also misfolds into tangles inside them. Both are hallmarks of the disease, detected decades before symptoms.
Alpha-synuclein misfolds and forms Lewy bodies inside dopaminergic neurons. Why it misfolds in some people and not others remains unknown. Genetic variants in the protein itself are one risk factor.
Prion proteins are infectious misfolded proteins — the only known biological entity that replicates with no DNA or RNA. They convert normal proteins into their misfolded shape on contact. There is no cure.
A single amino acid mutation (F508del) causes the CFTR protein to misfold and be destroyed before reaching the cell surface. No channel protein → no ion transport → thick mucus in lungs. The most common life-shortening mutation in people of European descent.
IAPP (islet amyloid polypeptide) misfolds and deposits in the pancreas, contributing to β-cell death. A different disease than Type 1, but protein aggregation is central to both.
p53 is the "guardian of the genome" — a protein that triggers cell death when DNA is damaged. 50% of all cancers carry a p53 mutation that causes it to misfold, lose its shape, and fail to function. Restoring its fold is an active drug target.
"If you know the sequence of a protein, you should be able to predict its structure — and from structure, its function."— Christian Anfinsen, Nobel Lecture, 1972
Frederick Sanger determines the complete amino acid sequence of insulin — the first protein ever sequenced. He wins the Nobel Prize. We can now read the letters. Predicting what they fold into is another matter entirely.
X-ray crystallography gives us the first 3D protein structure. Myoglobin, the oxygen-storage protein in muscle, is resolved at atomic resolution. The structure takes years of data collection and months of hand-calculation. Kendrew and Perutz share the Nobel Prize in Chemistry.
Christian Anfinsen demonstrates that a denatured protein can refold spontaneously into its correct shape. His conclusion: the native structure is entirely determined by the amino acid sequence. This is the founding axiom of the folding problem — and his Nobel Prize. It also means: if you could compute it, you could predict any protein structure from its gene.
Cyrus Levinthal calculates that if a protein sampled random conformations, it would take longer than the age of the universe. Yet proteins fold in microseconds. This "Levinthal Paradox" defines the problem: there must be a guided pathway, but we don't know what it is.
Critical Assessment of Structure Prediction: a biennial competition where teams predict protein structures from sequences alone, then real structures are revealed. The scoring system (GDT — Global Distance Test, 0–100) becomes the benchmark. Early winners score around 20–30. Human experts cluster around 90.
DeepMind enters a neural network trained on evolutionary sequence data. It wins the competition with a GDT score of ~61, far ahead of all traditional methods. The biology community notices. Two years of intense development follow.
At CASP14, AlphaFold 2 predicts structures with median GDT of 92.4 — exceeding human expert performance on many targets. John Moult, who co-founded CASP, says: "This is a watershed moment for biology." The 50-year-old problem is effectively solved. Science names it Breakthrough of the Year.
DeepMind and EMBL-EBI publish the AlphaFold 2 paper in Nature and simultaneously release the code. The architecture — attention-based neural network processing multiple sequence alignments into 3D geometry — becomes the most read structural biology paper ever.
The AlphaFold Protein Structure Database launches with predicted structures for nearly every known protein — including all 20,000 human proteins, all proteins in key model organisms, and UniProt's 200M+ entries. All free. All downloadable. The "dark proteome" — proteins never structurally characterised — is largely illuminated overnight.
AlphaFold 3 extends prediction to protein complexes, DNA, RNA, and small molecules — the full cast of biology. Drug discovery changes overnight. Hassabis and Jumper receive the Nobel Prize in Chemistry. The other half goes to David Baker for computational protein design — not just predicting natural proteins, but designing new ones from scratch.
The CASP competition ran every two years from 1994. For 24 years, the best human methods improved by fractions of a point per competition. Then in 2020, AlphaFold 2 scored 92.4 — leaping past human expert performance in a single jump. Hover each point to see the story.
Start with your protein sequence. Search all known sequences across evolution for similar proteins. Co-evolving positions — pairs of residues that mutate together — reveal which pairs are physically close in 3D space. This is evolutionary geometry.
A transformer-like attention mechanism processes every residue pair simultaneously, learning which positions constrain each other. Unlike older methods that looked locally, this sees the whole sequence globally at once. This is the key architectural insight.
Instead of predicting distances and then assembling a structure, AlphaFold 2 directly outputs 3D rotation and translation for each residue using equivariant neural networks. Iterative refinement cycles ("recycling") tighten the prediction.
For every residue, AlphaFold predicts its own confidence (pLDDT, 0–100). High confidence = reliable. Low confidence often means the region is intrinsically disordered in real life — which is itself biologically meaningful information.
Disordered regions often become structured only when they bind a partner. AlphaFold's low-confidence flag is not a failure — it's pointing at something biologically interesting.
The protein folding problem was not just an intellectual puzzle. It was the key blocking scientific progress in drug discovery, vaccine design, enzyme engineering, and our understanding of virtually every disease. Now that key turns.
Knowing a pathogen's protein structure immediately reveals binding pockets for drug molecules. AlphaFold structures have already been used to design drug candidates for malaria, tuberculosis, and antibiotic-resistant bacteria — in weeks rather than years.
PETase, a bacterial enzyme that breaks down plastic bottles, was redesigned using structural knowledge for 6× better efficiency. AlphaFold structures of related enzymes now accelerate this further. Enzymatic plastic recycling may be economically viable within a decade.
Designing vaccines requires knowing the exact shape of the antigen your immune system will train against. COVID mRNA vaccines were designed using knowledge of the spike protein structure. Future vaccines for RSV, HIV, and cancer will be faster to design.
~138,000 people die from snakebite annually, mostly in the Global South. Antivenoms require milking live snakes and are expensive to produce. AlphaFold structures of venom proteins are enabling rational design of cheap, synthetic, broadly effective antidotes.
AlphaFold's inverse problem: given a desired function, design a sequence that folds into a shape that achieves it. Baker's lab (Nobel 2024) has designed proteins that don't exist in nature — vaccines, nanomachines, self-assembling materials — purely from first principles.
Before AlphaFold, roughly 1/3 of human proteins had no known structure. These "dark proteome" proteins were often the most interesting — disease-relevant, drug targets, but too difficult to crystallise. AlphaFold predicted nearly all of them. Thousands of papers are being written about proteins that were structurally invisible last year.
"AlphaFold is a once in a generation advance — one of the most important scientific achievements in decades."— Edith Heard, EMBL Director-General, 2021
The protein folding problem seemed intractable because we tried to solve it as a physics problem — simulate every atom. AlphaFold solved it as a pattern recognition problem: 200 million years of evolution had already run the experiment. The answers were hidden in the sequences all along. We just needed to learn to read them.
Evolution is the world's longest running simulation. AlphaFold learned to play it back.