Demis Hassabis, Father of Google's Artificial Intelligence

Part 6. Digital Biology

15 The Protein Folding Problem

Kim Kyung-ran, Kim Kyung-jin

In December 1972, the winter in Stockholm was unusually cold, but the Nobel Prize ceremony hall was charged with excitement. Christian Anfinsen, an American biochemist, stepped onto the stage and announced a bold hypothesis that would mark its place in the history of science. He presented a principle about how proteins, the microscopic machines of life, are made. It was strikingly simple yet fundamental.

"The three-dimensional structure of a protein is determined entirely by its amino acid sequence." This is the proposition that would later be known as 'Anfinsen's Dogma.' The declaration handed the world of biology an enormous challenge.

If Anfinsen was right, then in theory, knowing only the one-dimensional amino acid sequence should let us perfectly predict the three-dimensional shape it would twist and fold into. But nature was not so obliging. The number of possible ways an amino acid chain could fold exceeded the total number of atoms in the universe.

A scientist named Cyrus Levinthal ran the numbers and raised what became known as 'Levinthal's Paradox': for a single protein to randomly try every possible structure would take longer than the age of the universe. Nature finishes the process in a thousandth of a second, yet human supercomputers could calculate for decades and still fail to find the answer. To tackle this problem, scientists launched a very unusual experiment in 1994.

It was called CASP, the Critical Assessment of Structure Prediction. Founded by Professor John Moult of the University of Maryland, the competition was a kind of blind test. Organizers would pose protein problems whose structures had already been solved in the lab but not yet published. Computer scientists and biologists from around the world would then submit their algorithmic predictions. Held every two years, the competition was the Olympics of structural biology.

As the competitions went on, the mood among scientists grew darker. From 1994 through 2016, over roughly twenty years, prediction

accuracy barely budged. Scores hovering around 40 out of 100 seemed to expose a hard ceiling on human intellect. Scientists were slowly wearing down.

Then, from a research lab in London, a man appeared who saw the problem from an entirely different angle. Why protein folding is a 'root node' problem: the bottleneck of life science. To Demis Hassabis, the world was one vast information-processing system. When he set DeepMind's mission as 'Solve Intelligence,' he was not dreaming of an AI that merely played Go well or chatted fluently.

He wanted a meta-solution, one that could dramatically accelerate the pace of scientific discovery. Hassabis compared science's many hard problems to tree branches. Climate change, disease treatment, new materials development: these were leaves and fruit hanging at the tips of branches.

Solving each of those problems one by one would cost enormous time and money. Hassabis thought: 'Instead of trimming branches one at a time, what if I found and solved the root node, the single point from which all these problems grow?' In his view, biology's root node was protein folding.

Every function in the human body runs on proteins. Sensing light with the eyes, moving muscles, antibodies fighting viruses: all proteins. A protein's function comes from its shape. Just as a key must fit a lock precisely to open a door, a protein needs a specific three-dimensional structure to unlock the lock of disease.

Drug development is so difficult and expensive because scientists did not know the exact shape of the target protein a drug needed to act on, forcing tens of thousands of rounds of trial and error. Hassabis was convinced: if AI could predict protein structures, it would be the catalyst that transforms biology from a purely experimental science into a data science. This was the perfect challenge to prove that DeepMind was not a game company but a scientific research organization.

From games to science: the inspiration drawn from the game 'Foldit.' The decisive moment that cemented Hassabis's confidence in the protein problem came, surprisingly, from a game.

In 2008, a team led by Professor David Baker at the University of Washington developed a game called Foldit. It displayed complex protein structures on screen as three-dimensional puzzles, and ordinary gamers used their mice to twist and fold them into the most stable configuration. What happened next was astonishing.

Gamers playing Foldit solved the structure of an AIDS-related enzyme in just three weeks, a problem that scientists had failed to crack despite running supercomputers for over a decade. These players had no Ph.D. in biology. They were simply people with a talent for recognizing patterns in three-dimensional space and fitting puzzles together by intuition. When Hassabis heard this news, a flash of insight struck.

'If human gamers use intuition to narrow a vast search space, isn't that exactly what AlphaGo did?' Go also has more possible positions than atoms in the universe, but instead of calculating every move, AlphaGo found the path to victory through its 'value network' and 'policy network,' a form of trained intuition. For Hassabis, protein folding was no longer a biology problem.

It was a spatial optimization game and a pattern recognition problem. Even without biological expertise, as long as there were data and a reward signal, AI could perform far better than any human gamer. To him, a former game developer, science's great challenges began to look like boss monsters waiting to be beaten.

The launch of the AlphaFold project: 'This is the lighthouse project.' In 2016, right after AlphaGo defeated Lee Sedol 9-dan, an intense debate erupted inside DeepMind: 'What comes next?' Some argued for conquering more complex games like StarCraft. Others pushed for robotics. Hassabis, quietly but firmly, began assembling a science team.

He named it the 'Lighthouse Project.' Just as a lighthouse shines its beam for ships lost at sea, the project was meant to serve as proof that AI could deliver real, tangible help to humanity.

The start was far from smooth. Even core DeepMind engineers pushed back: 'Biology is too uncertain,' 'The data is messy and hard for AI to learn from.' Hassabis rallied the team with his characteristic persuasion.

He brought together a small, elite group including John Jumper, a young researcher in his late twenties, and formed the AlphaFold team. The goal was clear: 'Take the victory won on the Go board and carry it onto the canvas of life.'

Hassabis sensed that if this project failed, DeepMind would be remembered as nothing more than a company that built machines good at games. If it succeeded, it would be a Nobel Prize-caliber discovery. It was the biggest bet placed simultaneously by Hassabis the scientist and Hassabis the executive. The three-dimensional folding structure of proteins.