1. In a Given Environment
A protein is a long chain made by joining twenty kinds of amino acids, one after another. Short chains have a few dozen links, long ones a thousand or more. In a particular environment many chains fold into fairly stable three-dimensional shapes, though some parts don't have a fixed shape. Where a protein can grab another molecule or cut a bond has to do with its shape.
On December 11, 1972, Christian Anfinsen gave his Nobel lecture in Stockholm. He had spent years experimenting with an enzyme, undoing its shape and putting it back into a suitable solution, where it folded back into its original shape on its own. In the lecture he called the conclusion the thermodynamic hypothesis. Its central claim was that the native conformation is determined by all the interactions between the atoms, and thus by the amino-acid sequence, in a given environment. In parentheses beforehand he listed some conditions: solvent, pH, ionic strength, metal ions, and temperature.
What traveled was that sequence determines structure. In a given environment didn't make the trip. Neither did the list in parentheses.
2. How to Fold Graciously
In 1969 the University of Illinois held a conference on a kind of spectroscopy in biological systems. The molecular biologist Cyrus Levinthal gave a talk three pages long called “How to Fold Graciously.”
He ran the numbers. A protein chain could take on something like ten to the three hundredth possible shapes. There was no way to try them one by one. Yet a natural protein could fold in seconds. His explanation was that the rapid formation of local interactions speeds and guides protein folding, and that those interactions then determine how the rest of the chain folds. It doesn't try every shape. It makes a few connections in small places, and the rest follows.
At the end of the talk he showed a short film. On the screen of a computer-controlled oscilloscope, a polypeptide chain grew little by little, taking an energetically favorable path, making connections where they belonged.
The word paradox isn't in those three pages. In 1992 three physicists published a paper called “Levinthal's Paradox.” A man gave a talk on how to fold graciously. Twenty-three years later his name had a paradox after it.
3. Seven Structures
The traditional way to find a protein's shape was to make it, purify it, grow a crystal, shine radiation on it, and work backward from the resulting spots in the diffraction pattern. Nuclear magnetic resonance and cryo-electron microscopy came along later. A single structure often took a few people months, sometimes years. Working out a structure by hand from nuclear magnetic resonance measurements alone generally took trained specialists weeks to months.
In 1971 there was a meeting at Cold Spring Harbor, on New York's Long Island. Thirty-five years later Louise Johnson recalled that Max Perutz had brought together people who already had atomic coordinates for proteins. She named several, including herself. Perutz wanted other people to be able to get hold of those coordinates easily, and he wanted them to be kept for the long term. On October 20 that year a journal in the Nature family printed a short announcement: Protein Data Bank. The bank was based at Brookhaven National Laboratory and started out with seven structures.
From then on, a lab anywhere in the world could determine a structure and put it in. In 1989 a committee of the International Union of Crystallography required papers on protein crystal structures to deposit their coordinates in the bank. In 1999 the US National Institutes of Health required work it funded to make the coordinates public when the paper appeared. By 2012 the bank held more than eighty thousand structures. It passed a hundred thousand in 2014 and a hundred and fifty thousand in 2019.
Each one took a group of people months or years.
4. Players
In 2008 researchers at the University of Washington released a computer game called Foldit. Players dragged a protein chain around the screen, bending and twisting it to make its shape use less energy. The system gave them a score.
In September 2011 a structural biology journal in the Nature family published “Crystal Structure of a Monomeric Retroviral Protease Solved by Protein Folding Game Players.” The protein came from a monkey virus. The crystal and its diffraction data were already there, but the structure had resisted being worked out. The researchers made it the subject of a three-week contest in the game. Some of the players' models matched the diffraction data closely enough for crystallographers to work with. Then the crystallographers completed and refined the structure. A headline at the time said players had solved in ten days a structure that had eluded scientists for fifteen years.
The third and fourth entries in the author list aren't individuals. They're two teams of players: Foldit Contenders Group and Foldit Void Crushers Group.
5. The Test
In 1994, John Moult and three colleagues organized the first assessment of protein structure prediction, called CASP. About thirty-five groups submitted a hundred and thirty-five predictions. The assessment was held every two years after that.
Here is how it worked. The organizers looked for labs that were close to determining a structure but hadn't made it public. They asked those labs to send their answers in advance and kept them under seal. They gave the prediction teams only the protein's amino-acid sequence and a deadline for turning in a prediction of its shape. Once a lab made its structure public, the predictions could be checked against it. The labs set the questions. Each question was the work of months or years. They held the answers and didn't say them yet.
One of the most widely used comparisons came to be the global distance test. Lay the predicted shape over the one determined by experiment and see what fraction of the positions along the chain falls within one, two, four, and eight ångströms of the experimental positions. Take the average of the four proportions for a score out of a hundred. An ångström is a ten-billionth of a meter.
6. In a Sense
December 2018, Cancún, Mexico. At the thirteenth CASP, DeepMind, the London company in essay seventeen, entered for the first time with a program called AlphaFold. In the hardest category, for structures with no close known match to refer to, it came first with a score of 68.3. The runner-up scored 48.2.
A university researcher who studied structure prediction, Mohammed AlQuraishi, went back and wrote a blog post. He opened with something someone there had asked him: what just happened? The advance, he wrote, was about what they'd expect from two rounds of CASP. Most of the academic researchers he'd spoken to felt something close to existential anxiety. Dozens of academic groups, a few hundred people in all, worked on this problem. DeepMind had entered a team of about ten.
In 2020 the fourteenth assessment was held online because of the pandemic. The results were released on November 30. On about two-thirds of the targets AlphaFold scored above ninety, roughly the range of accuracy of experiment itself. The median AlQuraishi gives rose from 68.5 in the previous round to 92.4. This AlphaFold used a new architecture.
That day a reporter from Science quoted CASP cofounder Moult calling it a big deal and saying that in a sense the problem was solved. CASP's own press release said that for individual proteins the problem was largely solved. DeepMind's blog ran under the headline “AlphaFold: a solution to a 50-year-old grand challenge in biology.”
Two days later the British structural biologist Stephen Curry wrote a blog post titled “No, DeepMind has not solved protein folding.” He distinguished predicting the shape at the end from finding out how a chain actually folds its way there.
7. Two Hundred Million
On July 15, 2021, “Highly Accurate Protein Structure Prediction with AlphaFold” appeared in Nature, with thirty-four authors. The code was released at the same time. That month David Baker's lab at the University of Washington published its own RoseTTAFold work.
On July 22 DeepMind and the European Bioinformatics Institute launched a database of predicted structures. The first batch had more than 350,000, about twenty thousand of them for human proteins. In July 2022 it grew to more than two hundred million. The institute's announcement said this covered virtually every catalogued protein known to science. Catalogued meant that the sequences were on record. The two hundred million and more shapes in the database had been calculated, not measured.
Every position along a predicted chain in the database gets a confidence score. Above ninety is very high. Below fifty is very low. The documentation says a very low score may correspond to a stretch that has no fixed shape to begin with, but the score alone cannot tell you that.
Later research tested the boundary one example at a time. Some proteins have more than one shape and shift between two or three of them; by default the system returns just one. Some proteins change shape only when they bind a small molecule; the system doesn't know whether that molecule is present. Change one amino acid, and the changes in shape the system predicts scarcely correspond to the changes in stability measured in experiments.
In 2012 Ken Dill and Justin MacCallum wrote a review in Science called “The Protein-Folding Problem, 50 Years On.” The fifty years ran from the 1962 Nobel Prize for protein structures awarded to Perutz and his colleague. They divided the problem into three: the physical folding code, how a chain folds over time, and whether a computer can predict a structure from its sequence. At the end they said it was no longer useful to speak of solving protein folding as a single goal.
The title of the AlphaFold paper uses the last of those three: structure prediction.
8. Folding
On October 9, 2024, half the Nobel Prize in Chemistry went to David Baker for computational protein design. The other half went to Demis Hassabis and John Jumper of DeepMind for protein structure prediction. The day before, the physics prize had gone to Hopfield and to Hinton, whom we met in essay fourteen.
The AlphaFold 3 paper appeared in May of that year. The program could predict how proteins fit together with nucleic acids, small molecules, and ions. The code did not come out with the paper; an online service was made available instead. A group of researchers in structural and computational biology wrote an open letter to Nature saying this departed from the field's norms. In November the code was released for noncommercial use only.
Several things were cut along the way. Anfinsen's in a given environment. Moult's in a sense. Individual proteins in the CASP press release. Catalogued in the institute's announcement. Dill and MacCallum saying not to treat the problem as a single goal. With all of that gone, one sentence is left: a program solved the fifty-year-old problem of protein folding.
A name is left, too. The prize citation said structure prediction. What spread was folding.
AlphaFold learned from the structures in the database. From the first seven in 1971, for half a century, one lab after another, one group of people after another, grew crystals, shone radiation on them, worked back from the spots, solved the structures one by one, and deposited them. For a CASP test, a lab turned in a structure it had just solved but hadn't published. The answer was sealed, and others tried to predict it. In the author list of the Foldit paper are the names of two teams of players.
Anfinsen's given environment is outside the sequence. These people are outside the program. This series calls what remains outside, however much you calculate, the remainder.
The machines are still being built. Each generation says again where the boundary lies.