The Legible Fold

For fifty years a protein's shape was locked behind an experiment. AlphaFold made the whole protein universe legible, and, unusually for a machine, it ships a per-residue confidence that tells you exactly where to trust it. Rotate a real prediction of TP53 and watch its certain core give way to tails it honestly cannot place.

A protein is a chain of amino acids, and that chain is easy to read. What the protein actually does depends on how the chain folds into a three-dimensional shape, and that shape was the hard part. Christian Anfinsen had argued back in the 1970s that the sequence alone should determine the fold (Anfinsen, Science), which sounds like a recipe. It was not. A modest protein can access an astronomical number of conformations, yet it settles on one in microseconds, and predicting which one from the sequence resisted five decades of effort.

So we measured shapes the slow way, one at a time, with X-ray crystallography and cryo electron microscopy. Decades of that work filled the Protein Data Bank with on the order of 200,000 experimentally solved structures (RCSB PDB), while the number of known protein sequences ran into the hundreds of millions. Almost everything alive was spelled out and unseen.

Every two years the field runs a blind test called CASP, where groups predict structures that have been solved but not yet released. In 2020, CASP14, DeepMind's AlphaFold2 reached a median score of 92.4 GDT on a scale where a model close to an experiment's own error lands in the 90s, with a median backbone accuracy near 0.96 angstroms (Jumper et al., Nature). For hard targets the field had been stuck near 60 for years. The leap was about 32 points in a single competition (Google DeepMind).

The following year the method and code were published, along with a prediction of nearly the entire human proteome (Tunyasuvunakool et al., Nature). A fifty-year problem did not so much get chipped away as it fell over.

AlphaFold's insight is that evolution has already run the experiment millions of times. If two positions in a protein touch in space, a mutation at one is often compensated by a mutation at the other, so their changes correlate across related species. Line up enough relatives and those correlations sketch a contact map. The model turns that into coordinates, and then scores its own work.

1. 01

   Sequence

   One protein: a string of amino acids, and nothing else to start from.
   ›
2. 02

   Evolution

   Search millions of related sequences. Residues that mutate together across species tend to sit together in space.
   ›
3. 03

   Evoformer

   Attention blocks pass information back and forth between the aligned sequences and every pair of residues.
   ›
4. 04

   Structure

   Turn those abstract pairwise relationships into concrete 3D atomic coordinates.
   ›
5. 05

   Confidence

   Score every residue for how far to trust where it was placed. That score is pLDDT.

Below is AlphaFold's actual prediction of Cellular tumor antigen p53, the tumor suppressor mutated in roughly half of human cancers, pulled straight from the AlphaFold Protein Structure Database (AlphaFold Protein Structure Database, EMBL-EBI). Each segment traces the protein backbone from one amino acid to the next. By default it is colored by AlphaFold's own confidence. p53 is a good teacher because it is half and half: a compact DNA-binding core that AlphaFold places with high confidence, flanked by tails it honestly marks as unplaceable. Drag to rotate it, switch the coloring to read direction along the chain, jump between p53's structural domains, or scrub the slider to walk residue by residue and watch the confidence change under your hand.

Alpha-carbon backbone, 393 residues, model v6, mean pLDDT 75.1. Coordinates and per-residue confidence are CC-BY 4.0 from AlphaFold DB (AlphaFold Protein Structure Database, EMBL-EBI); the reading of them is mine.

The color you were looking at is a number the model outputs alongside every residue, called pLDDT, from 0 to 100 (AlphaFold Protein Structure Database, EMBL-EBI). It is not a decoration bolted on afterward. It is AlphaFold saying, position by position, how much to trust where it just placed an atom.

Look again at what happened when you scrubbed into the tails of TP53. The confidence collapsed, and that is correct. Those regions are intrinsically disordered: they have no single fixed shape to predict, and their low pLDDT is the model getting the answer right, not getting it wrong. The 2021 human-proteome paper showed exactly this, that low confidence tracks disorder rather than failure (Tunyasuvunakool et al., Nature). About 41% of this model carries that honest warning label.

And the confidence is not arbitrary. Line the domain strip up against the confidence plot and they turn out to be the same picture. The high-confidence core is exactly p53's DNA-binding domain (mean pLDDT 91), the folded part that actually reads DNA and where most cancer mutations land (UniProt Consortium), while the low-confidence stretches are the transactivation and regulatory tails that stay flexible on purpose. AlphaFold recovered the functional map of the protein and, in the same output, told you how sure it was of each piece.

That is the habit worth stealing. A great deal of the research on this site is about classification systems that make the world legible while quietly hiding the places they cannot see. AlphaFold is the rare instrument that renders something newly legible and, in the same breath, publishes the borders of its own legibility. It hands you the map and shades in the parts it is guessing at.

With EMBL-EBI, DeepMind turned the method into a public library. The AlphaFold Protein Structure Database grew from about a million predictions (Varadi et al., Nucleic Acids Research) to structure coverage for over 214,000,000 protein sequences (Varadi et al., Nucleic Acids Research), free and openly licensed. In 2024, AlphaFold 3 extended prediction from single chains to complexes with DNA, RNA, ions, and drug-like ligands (Abramson et al., Nature), and the Nobel Prize in Chemistry went to Demis Hassabis and John Jumper for structure prediction, alongside David Baker for protein design (The Royal Swedish Academy of Sciences).

The honesty cuts the other way too. A prediction is not an experiment. AlphaFold models a single static shape, it can be confidently placed and still be biologically incomplete, and structural biologists still validate the ones that matter in the lab. The confidence score is what makes that discipline possible: it tells you which predictions to lean on and which to treat as a hypothesis. The timeline, in five moves:

1. 1972
   Anfinsen's dogma

   Christian Anfinsen's Nobel lecture crystallizes the premise: a protein's amino-acid sequence should determine its three-dimensional shape. Anfinsen, Science
2. 2020
   AlphaFold2 wins CASP14

   At the field's blind test, AlphaFold2 reaches a median GDT of 92.4, roughly experimental accuracy, on problems where prediction had stalled near 60 for years. Google DeepMind
3. 2021
   The method, and the human proteome, go public

   The Nature paper releases the method and code; a companion paper predicts nearly the entire human proteome and shows low confidence tracks intrinsic disorder. Jumper et al., Nature
4. 2022
   The database opens the protein universe

   With EMBL-EBI, the AlphaFold Protein Structure Database expands from about a million predictions toward the whole cataloged protein universe, free and openly licensed. Varadi et al., Nucleic Acids Research
5. 2024
   AlphaFold 3 and a Nobel Prize

   AlphaFold 3 extends prediction from single chains to complexes with DNA, RNA, ions, and drug-like ligands; the Chemistry Nobel goes to Hassabis and Jumper for structure prediction and to Baker for protein design. Abramson et al., Nature

Every figure on this page re-derives from a small content-as-code dataset, and the demo is a real AlphaFold prediction parsed from the database entry for TP53(P04637). The dataset behind the page is a download:

1. Highly accurate protein structure prediction with AlphaFold. Jumper et al., Nature, 2021.
2. AlphaFold: a solution to a 50-year-old grand challenge in biology. Google DeepMind, 2020.
3. Principles that Govern the Folding of Protein Chains. Anfinsen, Science, 1973.
4. AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space. Varadi et al., Nucleic Acids Research, 2022.
5. AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences. Varadi et al., Nucleic Acids Research, 2024.
6. Highly accurate protein structure prediction for the human proteome. Tunyasuvunakool et al., Nature, 2021.
7. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Abramson et al., Nature, 2024.
8. The Nobel Prize in Chemistry 2024. The Royal Swedish Academy of Sciences, 2024.
9. AlphaFold DB entry: Cellular tumor antigen p53 (P04637). AlphaFold Protein Structure Database, EMBL-EBI, 2025.
10. AlphaFold DB: confidence (pLDDT) and how to interpret it. AlphaFold Protein Structure Database, EMBL-EBI, 2024.
11. RCSB Protein Data Bank. RCSB PDB, 2026.
12. UniProtKB: Cellular tumor antigen p53 (P04637). UniProt Consortium, 2026.

Predicted structures from AlphaFold DB are licensed CC-BY 4.0 by EMBL-EBI and Google DeepMind. Figures verified as of 2026-07-27. Next review 2026-10.