[AI Library] 16 The AlphaFold Revolution
Demis Hassabis, Father of Google's Artificial Intelligence
Part 6. Digital Biology
16 The AlphaFold Revolution
Kim Kyung-ran, Kim Kyung-jin
It was a late autumn day in 2020, when the world was still locked down by a pandemic. The scientific community had its eyes fixed on a Zoom screen. It was the announcement of results for the 14th CASP (Critical Assessment of Structure Prediction), an academic competition for protein structure prediction.
This competition, which began in 1994, is something like the Olympics of biology. It is a battleground where teams compete to see how accurately computers can predict the three-dimensional structures of proteins whose forms have not yet been determined by experiment. The organizer, Professor John Moult, put up a single graph on the screen. The graph showed a gentle upward slope representing decades of incremental progress by scientists, and at the very end, a single dot that shot vertically upward.
That dot was 'AlphaFold 2.' Professor Moult declared in a trembling voice: "This means that a 50-year-old problem has been solved."
This moment was not simply a case of a computer program beating humans. When AlphaGo defeated Lee Sedol 9-dan, the world felt surprise and fear at the same time, but when AlphaFold solved biology's grand challenge, scientists felt awe and relief. For Demis Hassabis, this moment was more important than the AlphaGo match. It was the instant his lifelong mission was proven.
The intelligence that had conquered games was finally beginning to unlock the secrets of life, the most complex problem in the real world. The prologue goes back two years, to the CASP13 competition in 2018. At that time, DeepMind appeared on this stage for the first time, carrying 'AlphaFold 1.' Back then, the biology community looked at these outsiders from Google's AI lab with skeptical eyes.
They assumed that computer scientists lacking biological knowledge could never grasp the complexity of proteins. But AlphaFold 1 took first place, scoring highest on 25 out of 43 problems. Compared to the second-place team, which scored highest on only 3 problems, the gap was overwhelming.
AlphaFold 1 used an approach that treated protein distances like pixels in an image. Just as computer vision recognizes a photo of a cat, it learned the distance map between amino acid sequences as if it were an image, then inferred the structure from that. But Hassabis was not satisfied.
They had won, but AlphaFold 1's accuracy still fell short of the level needed for use in a laboratory. Converted to a GDT (Global Distance Test) score, it was around the 60s. It could get the rough shape right, but the positions of individual atoms were often wrong. Hassabis gathered the team and said:
"We did not start this project to come in first place. We started it to solve the problem." He made the bold decision to scrap AlphaFold 1's code entirely and start over from scratch. At this point, the person who would change the project's fate entered the picture.
That person was John Jumper. Having studied chemistry and computer science at the University of Chicago, he was a researcher with deep expertise in protein simulation. After joining DeepMind, Jumper led the AlphaFold team, serving as the commander who translated Hassabis's vision into technical reality.
If Hassabis defined 'what' to solve, Jumper designed 'how' to solve it. The two built trust as scientific partners, going beyond a simple boss-subordinate relationship. Jumper pointed out the limitations of the CNN (convolutional neural network) image-recognition approach that AlphaFold 1 had used.
His insight was that proteins are not static images but three-dimensional graph structures whose parts constantly interact and move. In 2019, the lights in the DeepMind lab never went off. The team introduced a new AI technique called the 'attention' mechanism.
This was a technology primarily used in language translation. Jumper and his team treated amino acid sequences like sentences. Just as words in a sentence relate to each other depending on context, they made the AI learn on its own how distant amino acids pull and push against one another.
It was a moment where biological intuition and AI engineering came together perfectly. They named it the 'Evoformer.' Hassabis, watching this process unfold, described it this way: "We are not reading a biology textbook to the AI. We are letting the AI discover the principles of biology on its own."
The development process was far from smooth. Even with only a few months left before the CASP14 competition, the new model remained unstable. Team members were exhausted, and fear of failure crept in.
This is when Hassabis showed his distinctive leadership. He would visit the lab late at night, play chess with the researchers or share tea with them, tending to their mental state. He told them, "We are standing on the threshold of a discovery that humanity has waited decades for," lifting the team's morale.
Finally, just weeks before the competition, AlphaFold 2 began showing exponential performance gains. And the results of CASP14 in 2020 were staggering. AlphaFold 2 scored an average of 92.4 out of 100.
A score above 90 is generally considered identical to the actual structure within experimental error margins. In other words, when you overlay the structure predicted by AI with a structure analyzed over several years using equipment costing billions of won, such as X-ray crystallography or cryo-electron microscopy, the two are indistinguishable. Even in the 'free modeling' category, known to be the most difficult, AlphaFold 2 demonstrated overwhelming accuracy.
When the results were announced, the chat window and Twitter were flooded with astonishment from scientists around the world. Venki Ramakrishnan, a Nobel laureate and master of structural biology, confessed: "This is a problem I did not expect to be solved in my lifetime." Professor Andrei Lupas of Germany's Max Planck Institute said, "Structural biology will now be divided into before AlphaFold and after AlphaFold," calling it "a change like an earthquake."
This victory was not a mere technical achievement. It was the first powerful piece of evidence proving his lifelong philosophy: "Solve intelligence, and use it to solve everything else." It showed that AI trained in the virtual world of games could understand the physical laws and biological complexity of the real world and extend the limits of human intellect.
He watched the announcement via Zoom from his home in London, quietly smiling. The intellectual curiosity he felt looking at a chessboard at age four, the possibilities of simulation he dreamed of while developing games as a teenager, and all his efforts as an adult
to fuse neuroscience and AI converged into this single peak. The success of AlphaFold 2 became the decisive factor that led him and John Jumper to jointly receive the Nobel Prize in Chemistry in 2024. Hassabis believed that the future changes this tool would bring were a greater reward than the prize itself.
Humanity had now obtained the instruction manual for assembling the Lego blocks of life. The release of a database containing over 200 million protein structures. After AlphaFold 2's dominant victory at CASP14, Demis Hassabis's ambition did not stop at a competition trophy. He was preparing to act on the principle he had upheld since DeepMind's acquisition by Google: "The most powerful technology must be the most widely used."
There was a path where they could monopolize the technology and sell licenses to pharmaceutical companies or generate enormous revenue. Many investors and entrepreneurs expected exactly that model. But Hassabis and the DeepMind team chose differently. They decided to expand the horizon of humanity's biological knowledge in one leap.
In July 2021, DeepMind partnered with the European Molecular Biology Laboratory (EMBL) to release the 'AlphaFold Protein Structure Database' to the world, free of charge. It started with roughly 20,000 protein structures encoded in the human genome. Even that alone was an enormous volume, enough to require rewriting biology textbooks.
In the past, a single PhD student would spend four to five years determining the structure of one or two proteins to earn a doctoral degree. By comparison, AI had drawn the map of all human proteins in a matter of days. The real shock came a year later, in 2022. DeepMind updated the database and released the structures of nearly every known protein on Earth.
The number exceeded 200 million. It included the proteins of nearly every organism science has named: plants, bacteria, animals, fungi. Hassabis called it
"the Google Earth of biology." Just as Google Earth lets us peer into alleyways on the opposite side of the globe, scientists could now sit at their monitors, click a few times, and rotate in three dimensions the molecular machines at the deepest level of living things. The release of this database fundamentally changed the speed of scientific research.
It was as if scientists who had been exploring a cave relying on a single flashlight in the dark suddenly had the switch flipped to illuminate the entire cavern. More than two million researchers from 190 countries accessed the database. That number represents the majority of the world's biologists.
Researchers who had been wrestling with pipettes at the laboratory bench could now form hypotheses based on AlphaFold's predictions, design experiments, and dramatically reduce their probability of failure. Practical applications appeared immediately and explosively. One of the most dramatic cases was malaria vaccine research.
A research team at Oxford University had spent years trying to determine the protein structure of the malaria parasite but had been unable to make progress due to technical limitations. As soon as AlphaFold was released, they searched for the structure of the protein they had been studying and obtained in just minutes the answer they had been seeking for years. This allowed them to immediately move on to designing a new vaccine candidate.
This is a symbolic scene showing how AI gave scientists the gift of 'time,' the most precious resource. Plastic waste is one of humanity's headaches, and scientists at the University of Portsmouth were studying an enzyme (PETase) that breaks down plastic. Using AlphaFold, they traced this enzyme's evolutionary history and succeeded in redesigning its structure to break down plastic faster and more efficiently.
This opened the possibility of decomposing plastic that would otherwise take 500 years to degrade in just a matter of days, making it a concrete example of Hassabis's vision to "solve climate and environmental crises through AI." AlphaFold also became an essential tool in research on antibiotic resistance and rare genetic diseases. A team at the University of Colorado analyzed the structures of antibiotic-resistant bacteria to explore new treatments, and researchers in Norway used the data to understand the immune system of honeybees.
Even in agriculture, whether for conserving endangered species or developing crops resistant to climate change, AlphaFold's data became seeds that sprouted new discoveries. Hassabis defined the impact of releasing this massive dataset as the beginning of 'Digital Biology.' He believed that biology had now entered the domain of information science.
Life is ultimately a complex process of information handling, and AI is the most capable tool for interpreting that information. In one interview, he said: "When we released 200 million structures, it was a gift to humanity.
What matters more is that this is only the beginning. We now need to move beyond knowing the shapes of proteins to predicting how they interact with each other, how they become new drugs, and how they cure diseases." These achievements became the strongest basis for the 2024 Nobel Prize in Chemistry.
The Nobel Committee recognized AlphaFold not simply because it was technically excellent software, but because it accelerated scientific progress by providing researchers around the world with a powerful research tool. By sharing this tool rather than privatizing it, Hassabis and John Jumper realized science's inherent value of 'sharing and spreading knowledge' not in Silicon Valley's way, but in the most classical and academic way. "The Google Earth of biology" was the most grand answer to the question Hassabis had asked at age 12: "Can't this intelligence be used for something better?"
Beyond the 64 squares of a chessboard, it laid open to humanity's eyes the microscopic world of life, as vast as the universe itself. This was not the victory of a single genius, but a true triumph of collective intelligence, in which the tool that genius built enabled millions of scientists to win their own victories, each in their own place.
Comparison of AlphaFold predicted structure (blue) and experimental structure (green)
Kim Kyung-jin
Attorney · Former Member of the National Assembly · AI Policy Researcher
© 2026 Kim Kyung-jin. All rights reserved.