You are position 9 in a protein. A hundred thousand species have copied you, badly, for two billion years. Somewhere in those copying mistakes is the shape of you.
A protein starts life as a floppy string of beads. Seconds later it is a machine — a pump, a scissor, a light detector — because the string has folded into exactly one shape out of an unthinkable number of possible shapes. Nobody told it which. It just fell into the right one.
The bead you are touching is position 26, far away along the string but pressed against you in the folded shape. For two billion years, every time evolution mutated you into something with a different charge, position 26 had to change too — or the whole machine stopped working and its owner died. Neither of you knows this. Neither of you can see the other. But the record of your handshake has been written, over and over, into the genomes of every creature that inherited you.
This edition is about the moment researchers realised that the mistakes in evolution's copying — the typos — contain the folded shape of the protein, and that a machine could read it out. It is the idea that unlocked the most celebrated result in the history of applied AI: a fifty-year-old problem, declared solved on a Monday in November 2020.
In 1969, at a meeting in a house in Illinois, the molecular biologist Cyrus Levinthal set out a puzzle in a short paper with a wonderful title: How to Fold Graciously. Count the shapes a protein could take, he said. Every link in the chain can bend two ways, and each of those bends has, say, three comfortable settings. Multiply.
Turn the dial below and watch the number. Then watch how long it would take to try every shape, one every ten-trillionth of a second — faster than any molecule can actually move.
Longer than the universe has existed. By a lot.
This is Levinthal's paradox, and it is not really a paradox — it is a proof that folding cannot be a search. The protein is not trying shapes. It is falling downhill, fast, guided by local bits that snap together first. Levinthal said so himself in the same paper.
But the paradox left a much harder question sitting in the middle of biology for the next fifty years. If a chain of beads reliably becomes one particular shape, then the shape must be written in the beads. Christian Anfinsen had shown this in the lab: unfold a ribonuclease enzyme until it is a limp noodle, take the poison away, and it ties itself back into a working machine. He won a share of the 1972 Nobel Prize in Chemistry for it. The sequence contains the structure.
So: read the sequence, get the shape. Easy to say. For half a century, nobody could do it.
Here is the move that changed everything, and it is not a machine-learning move at all. It is a reading move.
Don't look at your protein. Look at everybody else's version of your protein — the same molecule in a mouse, a mushroom, a hot-spring bacterium, a sea urchin. Line them all up, one per row, letters in columns. That stack is called a multiple sequence alignment, and it is a laboratory notebook kept by evolution for two billion years, recording millions of experiments in what a protein can survive.
That is coevolution. Two positions that touch in the folded shape are locked together: a mutation that breaks the pairing is usually fatal, unless a second mutation repairs it. Positions that never touch drift independently. So columns that change together are probably touching — and touching is a fact about three-dimensional shape, extracted from nothing but text.
Below is a real, live-generated alignment of a small 32-bead protein invented for this page. Each new row is a new species. Press play and watch the wall fill.
Pick two columns. The grid on the left counts how often each pair of letters shows up together.
Mutual information is the honest name for "how much does knowing column i tell you about column j". Zero means nothing at all. It was first turned on protein alignments in 1994, in a paper by Ulrike Göbel, Chris Sander, Reinhard Schneider and Alfonso Valencia called Correlated mutations and residue contacts in proteins. The idea was right. The results were disappointing. And the reason why is one of the loveliest bugs in science.
Suppose A touches B, and B touches C. Then A and B move together, and B and C move together — and so, second-hand, A and C appear to move together too. A correlation counter cannot tell the difference between a handshake and a rumour. Every real contact spawns a crowd of phantoms, and after a while the phantoms outnumber the truth.
The fix arrived from statistical physics. Instead of scoring each pair on its own, build one model of the whole alignment at once, in which every pair has a coupling, and ask that model: what coupling does each pair need in order to explain all the correlations we see? Pairs that are only correlated second-hand get a coupling of nearly zero, because their correlation is already accounted for by the chain through the middle. Martin Weigt, Robert White, Hendrik Szurmant, James Hoch and Terence Hwa published this in PNAS in 2009; in 2011 Debora Marks, Lucy Colwell, Robert Sheridan, Thomas Hopf, Andrea Pagnani, Riccardo Zecchina and Chris Sander used it to build 15 real protein structures out of nothing but sequences.
Both maps below are computed live, in your browser, from the alignment you just collected. Flip between the 1994 method and the 2009 method and watch the fog burn off.
Collect some species first — with an empty alignment there is nothing to correlate.
Getting it right depends on how many relatives you can find
Share of the top 32 predicted pairs that really touch, measured live on this page as species arrive
Notice the shape of those curves. Both methods are useless with a handful of species. The correlation counter improves and then stalls, because more data does not cure a phantom. The global model keeps climbing. This is why the hunt for more sequences became a scientific project in its own right — in 2017 Sergey Ovchinnikov and colleagues went digging through metagenomes, DNA scooped straight out of soil and seawater from organisms nobody has ever cultured, purely to make alignments deeper.
A list of pairs that touch is not yet a shape. But it is very nearly one. Give a floppy chain a set of rules — neighbours stay 3.8 ångströms apart, these particular distant pairs must come within about 8, nothing may sit inside anything else — and then just let it fall. That is what the button does. The chain below starts as a random tangle and is pulled into shape by nothing but the pairs your alignment predicted.
Collect species, then press Fold it. Drag the picture to turn the molecule around.
Two things worth noticing, because both are real and both cost researchers years.
Too few species and it fails. Not "gets a bit worse" — fails, into a lump. Coevolution is a statistical signal, and statistics needs numbers. Proteins with no known relatives are still hard today.
Contacts cannot tell left from right. Press Mirror it. Every distance is identical; every contact is satisfied; the molecule is now its own reflection, and no list of pairs can object. Real proteins are emphatically handed — their helices spiral one way — so contact-based folding has always needed a separate hint to break the tie. There are published methods whose entire job is deciding which of the two mirror images is the real one.
Since 1994, protein prediction has had an exam with no way to cheat. It is called CASP — Critical Assessment of Structure Prediction — and it was founded by John Moult. Every two years, laboratories that have solved a structure in the lab, but not yet published it, hand over the sequences. Anyone may predict. The answers are already sitting in a drawer, unseen. Then they are compared, and the comparison is scored out of 100 on something called GDT — roughly, the percentage of the molecule you put in the right place.
Through the 2010s, coevolution methods pushed the field forward, and then deep learning was pointed at the same alignments. At CASP13 in 2018 a DeepMind entry called AlphaFold won — impressively, but within the ordinary range of academic progress. Two years later, the results of CASP14 were released on 30 November 2020.
CASP14, 30 November 2020: AlphaFold2's median score
GDT, out of 100. Around 90 is regarded as competitive with an experiment done in a laboratory.
A median of 92.4 across all targets. 87.0 in the free-modelling category — the hardest one, where nothing resembling the target has ever been seen before. The organisers' own threshold for "as good as doing the experiment" was around 90.
At the time, laboratories had painstakingly solved around 170,000 protein structures. The database of known protein sequences held about 180 million. In July 2022 the AlphaFold Protein Structure Database, built with EMBL-EBI, published predicted structures for over 200 million of them — very nearly every protein sequence known to science — free to anyone. In October 2024 the Nobel Prize in Chemistry went half to David Baker for computational protein design, and a quarter each to Demis Hassabis and John Jumper for protein structure prediction.
And the thread that runs all the way back to position 9? AlphaFold does not compute a covariance matrix and invert it, the way this page does. It reads the alignment with a neural network, end to end, learning for itself what to notice. But the reason an alignment is worth reading at all — the reason all that information is sitting there in the typos — is exactly the handshake between you and position 26.
A beautiful mental model that is wrong is worse than a plain one that is right. So here is what has been simplified, bent, or invented.
To make this work in a browser, sequences here are sampled from a statistical model whose couplings sit exactly on the true contacts — and the bench then infers those couplings back. Real evolution is nothing so tidy. Real relatives are also not independent: they share ancestors, so a thousand near-identical bacteria count for far less than a thousand species do here. Real pipelines down-weight similar sequences for exactly this reason. Contact recovery on this page is easier than it is in life.
Thirty-two beads, two helices packed side by side, a six-letter alphabet instead of the real twenty. The helix geometry is real (about 100° of turn and 1.5 ångströms of rise per residue; neighbouring backbone atoms 3.8 apart) but the molecule is not in any database. No real protein was harmed.
The bench inverts a covariance matrix — the mean-field approximation. The 2009 paper used message passing; later work used pseudo-likelihood, which is better and slower. All of them are doing the same thing: separating a direct coupling from a rumour.
It does not score pairs and then fold to satisfy them. It runs a deep network over the alignment and the emerging structure together, refining both, and it outputs coordinates directly with a per-atom confidence estimate. Coevolution is the physics of why the input is informative. It is not the method.
AlphaFold2 predicts one still photograph of a molecule that in life bends, breathes, binds partners and sometimes has floppy regions with no fixed shape at all. It is weaker where relatives are scarce. Predicting how a mutation changes stability, or what shape a protein takes when it grabs a drug, remains open. A great deal of biology did not end in November 2020.
The Levinthal figures use three settings per bond and a tenth of a picosecond per attempt — the standard textbook version of a back-of-envelope Levinthal himself wrote loosely. The folding distance is measured after best-fit superposition, and the page quietly reports whichever of the shape and its mirror image fits better, which real evaluations do not allow.
No marks are recorded anywhere. Wrong answers are the interesting ones.
—