LONDON
— Demis Hassabis has no interest in a laboratory.
The Nobel Prize-winning artificial intelligence pioneer is betting big on AI’s future in drug discovery. But he’s also grown exasperated at hearing other AI-focused biotechs talk about the primacy of the lab, which has been the backbone of drug discovery since pretty much forever.
As others call for more experiments and data from tinkering with drugs and cells, Hassabis sees excuses. They don’t need bigger labs. They need better ideas. And maybe better thinkers.
“I won’t name names of other AI biotechs — you know them very well — but I feel like they always talk about, ‘We just need more data,’ and ‘We don’t have enough data,’” Hassabis said. “It feels like a bit of a crutch. Like, make your algorithms better, your models better. You do have enough data — if you were innovative enough on your algorithm side.”
Hassabis, 48, is looking to prove that claim across two CEO roles. He not only leads Alphabet’s AI efforts as the head of Google DeepMind, but he also is in charge of the three-year-old biotech startup Isomorphic Labs, which believes AI will change how drugs are discovered.
Drugmakers have spent decades playing with computer models, trying to use technology to reduce the industry’s 90% failure rate for drugs. This time, they have massively more powerful computer chips and better algorithms. Billions of dollars in venture capital has poured in. And a new generation of leaders, like Hassabis, has arrived from the tech world, with radical ideas about how machines can do what people can’t.
Since launching in November 2021, Isomorphic has been one of the most discussed and debated startups in biotech. The company was launched to build on and beyond AlphaFold, DeepMind’s breakthrough protein-structure-predicting AI model now used across the industry.
AlphaFold opened many techies’ eyes to biology, and opened the wallets of venture capitalists hunting for the next big thing. Isomorphic has been at the forefront since, working in relative secrecy with occasional announcements of
multibillion-dollar pharma partnerships
or the
next generation of AlphaFold
.
Despite an explosion of new startups and AI models, Hassabis has his doubts if others are up to the task. How many of these other companies, he asks, could have created AlphaFold 2, the invention behind the recent Nobel Prize win?
“That’s the quality you require, which is a top-level machine learning team,” he said. “There’s only three or four of those teams in the world, and they’re mostly doing general AI currently. It’s quite hard to do that as a biotech.”
For the first time, Isomorphic opened its doors to a reporter, as
Endpoints News
exclusively toured its London headquarters last month. In sit-down interviews and a software demo that went beyond its so-far publicly discussed research, Isomorphic’s leadership gave a glimpse behind the curtain into how Hassabis’ AI-driven vision is becoming reality, and why the company believes its pure focus on technology and ideas will lead to success compared with companies taking different approaches.
There are, of course, plenty of doubters and cautionary tales of game-changing technologies that didn’t change the game. That includes much of
the first generation
of AI-driven biotechs that started a decade ago, and tech’s underwhelming forays into healthcare like IBM Watson and Alphabet’s Verily.
But Isomorphic’s team brings a unique blend of naivety and experience around Hassabis. There are longtime DeepMind leaders who likely haven’t touched a pipette since high school, but there are also biopharma veterans who have brought dozens of drugs into human testing. And as the startup enters its fourth year, its leaders say they are emboldened by the pace of progress on their laboratory-free, computer-heavy idea of what biotech’s future may look like.
“I’m very used to people saying things aren’t going to work,” Hassabis said, “and then, eventually, they do.”
For decades, biotech’s signature look has been scientists in white lab coats and safety glasses, toiling away at the lab bench. The industry is full of stories of researchers working around the clock to synthesize compounds, run experiments against them and then take the results to improve the next batch. It’s a process so fundamental to biotech that a key barometer of the industry’s health is how much new lab space is being built or rented in a given year.
Isomorphic is choosing a different way.
“It’s been intentional that we don’t want to get distracted by trying to set up our own labs or trying to have a lab-in-the-loop,” Isomorphic’s chief scientific officer Miles Congreve said.
Congreve balances out Hassabis’ dreams with a dose of hard-earned realism, having spent the last three decades in the drug industry at places like GSK and Sosei Heptares. As Isomorphic’s top scientist, he’s responsible for ensuring the no-lab mentality will, eventually, deliver real drugs.
The lab-in-the-loop philosophy has grown widespread, the phrase popularized by
Aviv Regev’s makeover of Genentech
but now embraced by drugmakers large and small. The goal is design-make-test cycles that improve the computational side by feeding back experimental results.
That may work for protein-based therapeutics, like cyclic peptides or antibodies, Congreve said, but Isomorphic is focused on small molecules. The chemistry needed to make molecules and explore the vast space of possible chemicals is simply not ready for a lab-in-the-loop, he said.
Instead, Isomorphic’s headquarters looks like any tech startup — rows of employees crouched over laptops or staring at desktop monitors, designing drugs on software. The biotech has built out a network of contract researchers, paying them to test their
in silico
creations in real life.
So while nearly all are betting on the lab-in-the-loop working, Isomorphic is pursuing the other extreme, hoping its computer-led process can reduce the number of design rounds needed for a typical drug project from 20 to a mere two or three, Congreve said.
“Our emphasis has been on reducing the number of molecules you actually need to make, rather than productionizing the experimental work,” Congreve said.
From the outside, it’s hard to say if Isomorphic’s strategy will be transformative, said Joshua Boger, the legendary founder and former CEO of Vertex Pharmaceuticals.
Boger is rooting for Isomorphic, calling the company a “first-class operation,” but is also skeptical that any single technology can upend drug R&D.
“Going to Mars is a chip shot compared to making a drug,” Boger said. “I just think they’re going to run up against the hard reality of drug development.”
The key challenge is the high clinical failure rate, Boger said. Changing that is a tall task, given the countless pitfalls that become apparent only through human testing, not in the lab or in computer simulations.
“Can they have a higher rate of coming up with things that look really good in cells? Maybe,” Boger said. “But I can tell you that is just not predictive of final success.”
Boger’s observation on the near impossibility of developing drugs is true. But from childhood, Hassabis has taken on increasingly difficult problems and bested them.
Growing up in northern London, he used his winnings from chess tournaments to buy his first computer at the age of 8. By 13, he was rated as a Master (a title held by a few thousand players at any given time, out of millions) and ranked second in the world for his age group. At 16, he was admitted into Cambridge University, but was too young to attend, so he worked at a computer-game company for a gap year. He was the lead programmer on the bestseller Theme Park at 17 years old.
Gaming and AI defined the majority of Hassabis’ early life, culminating in co-founding DeepMind Technologies in 2010. The goal was, and still is, to develop artificial general intelligence — AGI — or AI models that could effectively do tasks as well as, or better than, the human brain.
DeepMind first built AI models that mastered simple computer games like Space Invaders, eventually catching the eye of Google in 2014, when the tech giant acquired the startup for around $600 million. Operating independently from Google, Hassabis’ team continued on its gaming quest. More powerful models tackled more complicated games, culminating in a machine-versus-man showdown in Seoul, South Korea, over the board game Go.
Played on a bigger board with far more possible moves, Go is magnitudes more complex than chess for a computer (or human) to master. In challenging Lee Sedol, a world-class player, Hassabis’ confidence in AI was once again put to the test. Sedol predicted a landslide win, echoing most Go players’ expectations that Sedol would prevail.
For his successes, Hassabis has painfully miscalculated before. Before DeepMind, he tried creating video games far too complex and ambitious for their time — bets that racked up losses for his video game studio. With $1 million in prize money, and more importantly DeepMind’s reputation as an AI leader, on the line, Hassabis was betting the time was right to beat Sedol.
Sedol played five games against the AI model, AlphaGo. AlphaGo made one move in the second game that left the match’s spectators nearly speechless. One wondered if the human in charge of making AlphaGo’s moves may have misplaced the stone. Sedol methodically rocked in his chair, trying to figure out what was happening in a game he dedicated his life to mastering.
AlphaGo’s move broke with thousands of years of human experience, playing a move later estimated to have a 1-in-10,000 probability of being played at that moment. AlphaGo not only won Game 2 but trounced Sedol, beating him 4 games to 1.
On the flight back from Seoul, Hassabis huddled with David Silver, a longtime DeepMind colleague, about what’s next. They agreed the technology was ready to jump to the real world, with its next contest being against biology. The result was the Nobel-winning creation of AlphaFold. The AI model turned the sequence of a protein — a string of letters representing its amino acid makeup — into a prediction of its three-dimensional shape.
The interest in biology didn’t come out of the blue. Before starting DeepMind, Hassabis earned a PhD in cognitive neuroscience, partly in the belief that studying the brain could help build intelligent machines. But Hassabis said he’s long viewed biology as a top application for AI, once powerful enough. AlphaFold was a test to see if the time was right.
The first version of AlphaFold achieved roughly 60% accuracy in a structure-predicting contest in 2018 — enough to win, but not enough to solve the problem. Human-led efforts had plateaued for years at around 40% accuracy, by comparison. Unsatisfied with 60%, DeepMind rebuilt the system from the ground up into AlphaFold 2, a model that won the next competition in 2020 with an average accuracy of nearly 90 on that same 100-point scale. No other team was close.
That was an astonishing accomplishment, one that turned what could be a year or two of work for a PhD student into a nearly instant computation. While Hassabis and his team relaxed over Christmas break, AlphaFold 2 ran continuously, predicting the structures of nearly all 20,000 proteins expressed in humans.
Hassabis then took the logical next step: Maybe AI was ready to upend biotech.
Isomorphic officially launched in November 2021, seeking to remake drug R&D using AI at the core, not just as an aid. It started as nothing more than a pitch deck that Hassabis and Colin Murdoch, a longtime DeepMind sidekick who’s now Isomorphic’s president, presented to Alphabet executives on a video call.
In its first months, the startup was just 10 people, who started to draft what came next. From the start, it was evident AlphaFold wasn’t going to move the needle on its own. They mapped out the arduous process of drug discovery on a whiteboard in their new headquarters, directly next door to DeepMind’s office, identifying a half-dozen or more tasks that required AlphaFold-like breakthroughs, from predicting how proteins and other biomolecules interact to the drug-like properties that distinguish viable molecules from toxic chemicals.
Hassabis said one of the biggest misconceptions of Iso’s work is that people think they’re trying to change R&D with just AlphaFold, which is now in its third generation, AlphaFold 3.
"The claim is not, ‘AlphaFold 3 solves drug discovery,’" Hassabis said. "It's just one piece of the puzzle. An impressive piece, a necessary piece, but it's not enough on its own, clearly. And no one's ever said it would be."
Today, Isomorphic has grown to about 150 employees and in May 2023 opened a second office in Switzerland. On a crisp Monday afternoon in November, about 15 of them convened around a U-shaped arrangement of tables in their company cafe to share a demo of their software with Endpoints.
The chemists at Isomorphic design and test molecules from their laptops. The example included a table with over 100 rows, each for a molecule. There are over a dozen columns, each carrying a property prediction for that molecule. When a chemist clicks on one of the rows, a 3D visualization appears, showing how the molecule may bind to a target protein (powered by what the team called “beyond AlphaFold 3,” its latest, not-yet-public model.)
Other columns feature predictions using different AI models than AlphaFold, like in assessing such drug-like characteristics as lipophilicity and solubility. There’s a synthesizability score, from 0 to 1, that ballparks how easy (or hard) it is to make that compound in the real world.
The demo gets more intriguing when a chemist starts playing with the software. In one mode, called guided design, they edit a molecule — drawing on or stripping away chemical parts. The dozen-plus predictions update automatically, giving instant feedback on design changes.
In another mode, called unguided design, the chemist no longer tries to tinker their way to a better molecule. Instead, they input some goals, like a certain binding strength and a minimum synthesizability score. The computer then explores chemical space for qualifying molecules, which the team say go far beyond the conventional wisdom of human chemists.
"Historically, chemists would make quite small changes to eke out some advantage in binding affinity," Isomorphic’s chief technology officer Sergei Yakneen said. "They wouldn't want to mess with that molecule too much, because that might really mess up that binding affinity."
As demos tend to go, this ends in success, generating a slew of new molecules with the desired marks, ready for further research. Just as predictably, the example is unsatisfying in the lack of specifics. The idea of
in silico
property predictions is far from novel: The significance depends on just how accurate Isomorphic’s work is — if it can reliably predict the real world.
On that front, Isomorphic has not published on the performance of many of the AI models outside of AlphaFold that power its software. Still, Iso’s executives repeatedly said they’ve made far more progress than they would have anticipated by this point, and the software predictions are translating to the real world.
“We can’t be kidding ourselves that we’re getting great numbers from models on benchmarks if it doesn’t translate into real drug design success,” Isomorphic’s chief AI officer Max Jaderberg said.
If Isomorphic’s team can truly rely on these predictions, that would open up new ways of finding molecules. Conventionally, each molecule has to be synthesized, with each property prediction requiring its own experiment to measure.
“You’re talking about spending weeks or months waiting for the results of determining all of these individually,” Yakneen said. Isomorphic’s approach “allows you to explore a massive number of different hypotheses, which you otherwise wouldn’t be able to do because you’d just be waiting forever,” he added.
Isomorphic has brought in veteran drug developers under the longtime industry leader Congreve, including former research scientists of Eli Lilly and Pfizer, as well as veterans of earlier AI-focused biotechs like Exscientia and BenevolentAI.
“I don’t understand the first thing about how the technology is actually working, but I can understand what it’s doing and how we can best apply it in the context of our projects,” Congreve said.
Perhaps the critical link, one missing in many AI-focused biotechs, is building an equally large product team to connect Congreve’s drug designers and Jaderberg’s AI engineers. They give the software, now in its seventh version with weekly updates happening, a Google-y accessibility so veteran drug hunters like Congreve can put it to use.
It’s unclear how much longer Isomorphic will have before facing the typical biotech pressures of hitting milestones like getting into human testing and delivering study readouts. Hassabis declined to provide an exact timeline or further details on its pipeline, beyond a focus on cancer and immunology. He said they could “start thinking about getting into the clinic” by the end of next year.
Hassabis envisions a potential business worth upwards of $100 billion. Unlocking that value is a trickier task. Generally, biotechs have two business models: 1) Develop drugs and sell them, either to patients or to a larger drugmaker; or 2) sell software and services to pharma companies.
Isomorphic is going the first route, planning to sell off internal drug candidates before they make it to market.
“We felt that the right way to do this initially was to sell molecules,” said Murdoch, Isomorphic’s president. “You just don’t capture the value from selling the software.”
In a way, DeepMind’s victory over a human Go master may offer a preview of what’s to come in biotech, if Isomorphic is successful.
Three years after losing to AlphaGo, Sedol retired from professional play. “Losing to AI, in a sense, meant my entire world was collapsing,” he
said earlier this year
. “I could no longer enjoy the game.”
Drug discovery is far more complicated than a game like Go. But if Isomorphic is right, the efforts of human drug hunters may look prehistoric in hindsight, trying to tackle problems fundamentally better addressed by machines. Even if that means longtime developers like Congreve follow in Sedol’s path.
“Eventually, we should have a product that actually does a lot of that thinking for you,” Congreve said. “It becomes much less of a technical exercise and people like me can retire and just do it all
in silico
.”
