Demis Hassabis, Father of Google's Artificial Intelligence

Part 6. Digital Biology

18 The Dawn of Digital Biology

Kim Kyung-ran, Kim Kyung-jin

A new horizon in structure prediction. In May 2024, Google DeepMind's headquarters near King's Cross in London was once again gripped by quiet excitement. Just four years earlier, they had stunned the world by solving the fifty-year-old grand challenge of protein structure prediction with AlphaFold2. But for Hassabis and his team, that was not an ending but a beginning. If AlphaFold2 had revealed the shape of proteins as solitary sculptures, now they needed to uncover how those sculptures join hands and dance together.

Because life is not a static structure but interaction itself. Hassabis had always told his staff, "Biology is not a static noun but a dynamic verb." The crystallization of that philosophy was AlphaFold3.

The arrival of AlphaFold3 was not a mere performance upgrade over the previous version. It was a fundamental redesign of architecture and an expansion in how biology is viewed. Where AlphaFold2 focused on taking amino acid sequences as input and drawing out a protein's three-dimensional structure, AlphaFold3 listened to the conversations among nearly all molecules that constitute life.

Proteins in our bodies do not exist alone. They cling to DNA, the blueprint, and read out genes. They bind with RNA to relay information. They endlessly attach to and detach from small signaling molecules called ligands. Ligands in particular are the master key to drug development.

Most drugs work by finding a ligand, a key that fits precisely into the lock of a protein. AlphaFold2 showed the shape of the lock but could not perfectly show what it looks like with the key inserted. AlphaFold3 sparked a revolution at exactly this point.

What makes this technically fascinating is that Hassabis and the research team introduced a diffusion model. The technique is similar to what AI systems like Midjourney or DALL-E use when they generate images from text input. Starting from a blurry, noise-filled cloud-like

state, the model progressively removes noise until it produces a sharp, accurate molecular structure. Hassabis often compared this process to "truth gradually emerging from fog." By adopting this diffusion model, AlphaFold3 partially replaced or supplemented the earlier Evoformer module, enabling it to handle not only proteins but also DNA, RNA, various ions, and modified molecules within a single unified network.

This meant that AI had moved beyond biological intuition, drawing closer to the probabilistic essence of physical binding. The achievement sent an immediate shockwave through the scientific community. Structures of massive complexes that previously required months or years to confirm using X-ray crystallography or cryo-electron microscopy could now be simulated in just minutes.

For example, if you can determine the structure of a protein that binds to a specific region of DNA to help cancer cells proliferate, designing a drug to block that binding becomes far easier. AlphaFold3 improved accuracy in predicting protein-DNA interactions by more than 50 percent over the best existing tools. The thrill Hassabis felt as a teenager reading moves ahead on the chessboard, the joy of discerning invisible patterns, was now being recreated at the most microscopic scale of life. He published the paper in Nature and decided to make the tool freely available to scientists worldwide (for non-commercial use), an act that once again proved his longstanding belief: "The most powerful tools are most valuable when they are most widely used."

AlphaFold-Multimer and protein complex research. On the road to AlphaFold3, there was an important stepping stone: AlphaFold-Multimer, released in late 2021. This stage marked the decisive moment when Hassabis's DeepMind team broadened its view from individual entities to systems.

Take hemoglobin, a textbook example. Hemoglobin is not a single lump of protein but a complex of four subunits joined together with exquisite precision. If these four pieces do not interlock perfectly, our bodies cannot carry oxygen.

If AlphaFold2 masterfully sculpted the shape of each piece, AlphaFold-Multimer was the solver that figured out how they assemble. John Jumper, who served as a lead researcher on the team, used to recall those days with a quip: "We knew that proteins get lonely." In reality, nearly every function inside a cell is carried out by proteins pairing up or clustering together like enormous machines.

Signal transduction, molecular transport, cell division: all the core phenomena of life begin with binding. Hassabis sensed that predicting these complexes was not a matter of scaling up computation. When two proteins meet, subtle shape changes occur at their interface. Just as the shape of your hand shifts slightly during a handshake, proteins undergo a process called induced fit. AlphaFold-Multimer cross-analyzed multiple sequence alignment (MSA) data rich in evolutionary information and detected signals of co-evolution, traces of two proteins changing together over the course of evolution.

The logic is straightforward: if a specific site on Protein A changes and the adjoining site on Protein B changes in tandem, those two sites are likely part of the binding interface. The ripple effects of this technology were enormous. Structural biologists had long fought a puzzle-assembly battle to determine the structures of large protein complexes.

AlphaFold-Multimer was like showing them the completed picture on the puzzle box before they started. With it, scientists could simulate how the spike protein a novel virus uses to invade human cells binds to the human receptor, and they could peer at the molecular level into how immune cells recognize antigens. This went beyond satisfying academic curiosity; it provided a fundamental key to understanding the mechanisms of disease.

For Hassabis, the trajectory from AlphaFold-Multimer to AlphaFold3 connects to his AGI vision. He defines intelligence as "the ability to process information, predict the future, and plan." Biological phenomena can likewise be viewed as a process in which genetic information is translated into hardware called proteins, and these hardware units interact to run the software of life.

Therefore, when AI predicts these complex interactions, it means we have begun to understand the operating system of biology, that vast information-processing system. Hassabis often says, "We are applying intelligence trained in the digital world to biology, the most complex problem in the physical world." The AlphaFold series is the strongest evidence that his dream of 'AI for Science' is not a slogan but is actually expanding the horizons of human knowledge. On laboratory monitors, hundreds of thousands of atoms are now dancing, and humanity has at last begun to understand the choreography.

Isomorphic Labs spins off in 2021: turning drug discovery into a 'computable problem.' Accelerating drug development through AI: compressing ten years into months. In 2021, Demis Hassabis announced the founding of a new company under Alphabet, Google's parent. Its name was Isomorphic Labs. Drawn from the mathematical term isomorphism, meaning 'same form,' the name encapsulated Hassabis's deep insight: that biological processes and information-processing processes share a fundamentally identical structure. If DeepMind was the place that 'solved intelligence' in the domain of pure science, Isomorphic Labs was the operational unit that would use that solved intelligence to address humanity's most painful problem: disease.

Hassabis decided to retain his role as CEO of DeepMind while concurrently serving as CEO of Isomorphic Labs. This showed how much weight he placed on the project. Traditional drug development was akin to gambling premised on failure. Bringing a single new drug to market takes an average of more than ten years and costs exceeding two trillion won.

Out of tens of thousands of candidate compounds, researchers shuttle between beakers and petri dishes testing reactions one by one, proceed through animal trials, and even after entering clinical trials, the success rate remains below 10 percent. It was like searching for a needle in a haystack. Hassabis wanted to convert this inefficient process into a 'debuggable software problem.' The core strategy of Isomorphic Labs is to design small molecules that could become drugs, based on AlphaFold's predictive power.

Previously, determining the structure of a target protein alone took years. Now AI predicts the target's three-dimensional structure in days and simulates millions of compounds in virtual space to find one that fits snugly into the protein's binding pocket. This is a massive shift: replacing physical experiments with digital experiments (in silico). In early 2024, Isomorphic Labs signed strategic partnerships with global pharmaceutical companies Eli Lilly and Novartis, totaling three billion dollars (approximately four trillion won).

The fact that the conservative pharmaceutical industry placed an astronomical bet on a startup only three years old signaled that the possibilities AI had demonstrated had crossed from mere hope into conviction. Hassabis described the process by saying, "We are compressing time." In the past, chemists relied on intuition and experience to synthesize compounds. Now generative AI learns the laws of physics and biological constraints and proposes drug structures with architectures no human had imagined.

This is not only about going faster. It means exploring uncharted territories of chemical space that were never even attempted because of human bias. Researchers at Isomorphic Labs are now compressing the initial drug discovery phase, which used to take ten years, into months or even weeks. For patients suffering from disease, this means the speed of hope is accelerating.

The announcement of AI-designed drugs entering clinical trials in 2025. The year 2025 is poised to become a critical turning point in the history of digital biology. Hassabis and Isomorphic Labs designated 2025 as the year when drug candidates designed entirely by AI from scratch (de novo design) would enter clinical trials in earnest. While there had been earlier cases of AI being used to repurpose drugs or provide partial assistance, it is an entirely different matter for 'next-generation AI drugs' born from advanced structure-based drug design (SBDD) technology like AlphaFold3 to be administered to humans.

These drugs are targeting so-called 'undruggable' targets that were difficult to attack with conventional methods.

For example, cancer-causing proteins whose structures are too flexible for drugs to latch onto, or mutant viruses that have developed resistance to existing drugs. Isomorphic Labs used AI to predict even the dynamics of proteins, designing drugs that bind at the fleeting instant a protein assumes a particular conformation. The difference is as great as shooting at a still photograph versus shooting at a moving video. The announcement of clinical entry signals more than a technical achievement; it heralds changes in regulatory science.

Regulatory agencies such as the FDA are establishing new standards for evaluating the safety and efficacy of AI-designed drugs. Hassabis emphasizes throughout this process: "We are not just making drugs; we are reinventing how drugs are made." The 2025 clinical trials will serve as the ultimate proving ground to verify whether AI can move beyond being a lab assistant and become an autonomous inventor that saves lives. If these trials begin producing successful data, we will witness the full opening of the 'AI-native pharma' era.

The dream of virtual cell simulation. The ultimate destination that Isomorphic Labs and Hassabis look toward lies beyond drug development. It is the realization of a virtual cell. Current biological research resembles taking apart a complex clock piece by piece to understand its internals.

But knowing every part does not mean you perfectly understand how the clock works. A virtual cell would recreate the entire system of all those parts, proteins, DNA, metabolites, organelles, interacting inside a computer like a digital twin. This is the apex of 'simulation' that Hassabis has dreamed of since childhood. Just as he simulated the behavior of theme park visitors during his game development days, he now aims to simulate the behavior of molecules inside cells.

If a virtual cell is realized, we could administer a drug to the virtual cell before giving it to an actual patient and predict what side effects or chain reactions would occur. This would be an alternative that addresses the ethical problems of animal testing, and a method that could dramatically lower the failure rate of clinical trials.

Of course, this is a massive challenge that cannot be solved by AlphaFold alone. But Hassabis says, "We are now ready to put together the puzzle pieces of biology." AlphaFold predicts structures, other AI models predict metabolic pathways and gene regulatory networks, and these combine to form a grand 'digital organism.' That is the ultimate destination of the vision Isomorphic Labs declared when it spun off in 2021.

It may sound like a distant future, but just as AlphaGo conquered Go and AlphaFold conquered proteins, Hassabis's clock has always ticked faster than the world expected. The virtual cell is not a mere dream but a predetermined future growing clearer as data and computing power accumulate. Biology is an information-processing system. From research to product: the realities of regulation, clinical trials, and partnerships. "Biology is the most complex and beautiful information-processing system."

This declaration by Hassabis dragged biology from the domain of the wet lab into the domain of the dry lab, where data and algorithms rule. But turning a vision into a real product, an actual pill a patient can take, demanded a struggle of an entirely different dimension from pure scientific discovery. Even after reaching the pinnacle of scientific authority with the 2024 Nobel Prize, the complex realities of the pharmaceutical industry were massive barriers that could not be breached by AI model accuracy alone. The first practical challenge Isomorphic Labs faced was data asymmetry.

AlphaGo could generate infinite data through self-play within the perfect information environment of Go's rules. Drug development is different. High-quality biological data is mostly locked away deep in the vaults of major pharmaceutical companies, kept confidential. Hassabis's decision to join hands with traditional pharma giants like Novartis and Eli Lilly was not a choice but a necessity.

This was a strategy of combining DeepMind's brilliant algorithms (the engine) with the clinical data and chemical libraries (the fuel) that pharmaceutical companies had accumulated over decades. The partnership was like an encounter between two civilizations speaking different languages. Silicon Valley's emphasis on speed collided with the pharma industry's safety-first doctrine.

A software bug can be fixed with a patch, but a drug's side effect can threaten a human life. To bridge this gap, Hassabis recruited not only AI engineers but also seasoned medicinal chemists and biologists into Isomorphic Labs. "We are not an AI company; we must become an AI-native pharmaceutical company,

one that speaks AI as its mother tongue." That was his mandate. Regulation was another mountain to climb. Regulatory agencies including the FDA viewed AI's outputs as a black box.

"Why did the model predict this molecule would be effective?" Early deep learning models could not answer that question. Hassabis's team had to pour enormous effort into introducing explainable AI (XAI) technology to translate the basis of AI predictions into language that chemists could understand. This was the 'gate of trust' that laboratory innovations had to pass through before they could become hospital prescriptions.

Scientific accuracy versus business speed: the balance. Inside the person of Demis Hassabis, two selves have always coexisted: the scientist pursuing pure truth and the entrepreneur wanting to change the world. In the new territory of digital biology, these two selves engaged in constant dialogue and sometimes clashed. During the development of AlphaFold3 and the operation of Isomorphic Labs, this tension reached its peak.

Hassabis the scientist pursues perfection. He wants to understand nature's fundamental principles, to fully solve challenges like protein folding. For him, the measures of success were a Nature cover paper, victory at the CASP competition, and the Nobel Prize.

But Hassabis the CEO of Isomorphic Labs faces pressure to deliver 'good enough' solutions on time. In drug development, there are many moments when finding a candidate compound that can be quickly synthesized and is non-toxic matters more than a perfectly accurate structure prediction, even if it contains some margin of error. The balance was found in the design of a feedback loop. Isomorphic Labs is not a company that only runs AI models. They built their own laboratories to actually synthesize and test the predicted molecules.

AI designs, robots synthesize (Make), biological assays test, and the resulting data feeds back into AI learning: this is the DMTL cycle they built. Through this process, Hassabis learned how to increase business speed without abandoning his insistence on scientific accuracy. Rather than trying to predict everything, he chose a strategy of focus, concentrating computing resources on the critical interactions that decisively affect drug development.

This approach differentiated them from the hype pervading the AI industry. While many AI bio startups sold nothing but rosy futures, Hassabis emphasized scientific validation with almost cold-blooded rigor. "Biology doesn't lie."

He demanded not demos designed to impress investors but results reproducible in an actual laboratory. This patience was partly possible because of the protective shield of Google's vast capital, but it was equally driven by his firm belief that only scientific integrity ultimately guarantees commercial success. In the end, the vision of digital biology is not the arrogance of reducing life to zeros and ones in order to control it. It is the most concrete and noble enactment of the mission Hassabis has carried since age twelve, 'solve intelligence to save the world': remaining humble before the wondrous complexity of life, yet using intelligence, the finest tool we possess, to discern the patterns within that complexity and fix the bugs called disease to improve human lives. 3D rendering of a protein molecule