AlphaFold: A Protein’s Shape from Its Sequence

A Chain of Amino Acids

A protein is amino acids joined one after another. There are 20 kinds, with different side chains: some avoid water, some carry a charge, some are tiny. The brass strip at the front carries all 84 letters of this one, one letter for each amino acid.

The chain folds up by itself: the water-avoiding side chains tuck inside, the charged ones stay outside. The shape decides the job. CI2 comes from barley seeds; it puts out a loop that fits into the active pocket of subtilisin and chymotrypsin and blocks both enzymes.

In 1973 Anfinsen summed it up: in the right surroundings, the sequence alone decides the shape. Could the shape then be worked out from the sequence? The question stayed open for nearly fifty years.

CI2 · UniProt P01053 · 84 amino acids · one letter per amino acid

The Folding Problem

A chain can take more shapes than anyone can count. In 1969 Levinthal pointed out that a protein cannot try them one by one. As the arithmetic is usually told: even with only 3 shapes per amino acid, 84 of them make about 10⁴⁰ combinations; trying ten trillion a second would take about 10²⁰ years. CI2 folds in water in about 20 milliseconds.

So shapes were measured. In 1958 Kendrew used X-ray crystallography to get the first three-dimensional structure of a protein. Today the Protein Data Bank holds about 260,000 experimental structures, while 150 million protein sequences are known.

Since 1994 structure prediction has had its own contest, CASP: every two years, the targets are structures just measured and not yet released, and the entrants get only the sequences.

3⁸⁴ ≈ 10⁴⁰ shapes · PDB 260,896 experimental structures · UniProtKB 150 million sequences (2026-10-18)

The Clue Evolution Left

The same kind of protein turns up in many species, with different sequences but much the same shape. Lining them up position by position gives a multiple sequence alignment. The nine rows above are nine members of CI2’s family, from different species.

When two amino acids touch in space and one changes, the other often changes with it, so the two still fit. Finding pairs of positions that change together across thousands of sequences hints at which ones touch.

These numbers are the site’s own, from the family’s 3,163 sequences in Pfam: of the 32 most strongly co-varying pairs, 22 really are in contact in the crystal structure. The thin brass rods join those 32 pairs. AlphaFold, too, starts by reading an alignment like this.

Pfam PF00280 · 3,163 sequences · 22 of the top 32 pairs in contact (beta carbons under 8 Å apart)

A Note for Every Pair

The body of AlphaFold 2 is called the Evoformer. It keeps a set of numbers for every pair of amino acids: 84 amino acids make an 84 × 84 grid, the pair representation. The alignment keeps its own.

Over 48 blocks the two trade information: the co-variation in the alignment is written into the pair grid, and the pair grid steers how the alignment is read. Attention lets each cell look at the cells that bear on it. A triangle update adds one more rule: if i touches j and j touches k, the distance from i to k is not free.

The plate at the back shows the distance of every pair, measured in this prediction; darker is closer. The dark patches away from the diagonal are parts far apart along the chain that sit together in space.

84 × 84 pairs · 48 Evoformer blocks · plate: alpha-carbon distance of every pair in the prediction, 0 to 20 Å

Building the Third Dimension

The structure module turns the pair grid into coordinates. It treats each amino acid as a small rigid body, all starting stacked at the origin and facing the same way; over 8 layers each one moves into place and turns to its angle, and then the side-chain angles are worked out.

The whole network runs four times, each pass taking the last one’s result as input. This is called recycling.

The chain on the plinth is doing something else, computed live on this page: starting as a flat string, it moves step by step by gradient descent, using only the predicted distance of every pair, until the distances match. The first AlphaFold, in 2018, built structures from predicted distances in this way; the second put this step inside the network.

Computed on this page: gradient descent from a flat chain to the predicted shape; the readout is how far this step is from the prediction

How Sure It Is

AlphaFold gives every amino acid a score from 0 to 100, called pLDDT: how accurate it expects the local structure there to be. The database paints four colours: dark blue above 90, light blue from 70 to 90, yellow from 50 to 70, orange below 50.

Nearly all of CI2’s body is dark blue; the first 20 are yellow and orange. In the crystal this stretch cannot be seen either: the first 18 amino acids of the mature protein have no fixed place. Very low pLDDT often means a stretch is loose to begin with.

The plate now shows the predicted aligned error: lined up on amino acid i, how many ångströms amino acid j is expected to be off. Darker green means the two positions are more reliable relative to each other.

pLDDT ≥ 90 · 70–90 · 50–70 · < 50 · the first 20 are below 70

Against the Experiment

The steel wire is CI2’s structure measured by X-rays in 1987, PDB entry 2CI2. With the prediction laid over it, the 65 positions they share are on average 0.80 Å apart, less than the length of a carbon–carbon bond (about 1.5 Å). That number is computed on this page.

This is no exam, though: 2CI2 was released long ago and was in AlphaFold’s training data. The real exam was CASP14 in 2020, with unreleased targets. AlphaFold’s median backbone error was 0.96 Å, against 2.8 Å for the next best method; on CASP’s GDT_TS score (100 is a perfect match) its median over the domains was 92.4.

From 2021 the predictions went into the public AlphaFold Database. By 2024 more than 214 million protein sequences had a predicted structure.

Prediction against crystal: alpha-carbon RMSD 0.80 Å (residues 20–84) · CASP14: 0.96 Å against 2.8 Å

What It Cannot Do

A prediction gives one still shape. Proteins move and change shape as they work; AlphaFold usually gives one of those shapes, and says nothing about how the chain folds.

Proteins seldom work alone. AlphaFold 2 works on protein chains. AlphaFold 3, in 2024, takes proteins, DNA and RNA, small molecules, ions and modified residues together, and makes the atom coordinates directly with a diffusion module.

Its paper names the limits too: in loose regions it sometimes draws order that is not there, and its structures do not always respect chirality. A prediction is a hypothesis; before it is used for experiments or drugs, it often needs checking.

AlphaFold 3 (2024): proteins · nucleic acids · small molecules · ions · modified residues

The Nobel Prize

The 2024 Nobel Prize in Chemistry went half to Demis Hassabis and John Jumper, for protein structure prediction, and half to David Baker, for computational protein design: the other way round, choosing a shape first and working out a sequence that folds into it.

From Anfinsen’s 1973 statement that sequence decides shape to computing the shape from the sequence as an everyday tool took less than fifty years.

Nobel Prize in Chemistry 2024