The pioneer behind Google Gemini is tackling a fair greater problem—utilizing AI to ‘remedy’ illness

The astonishingly prodigious little one of bohemian mother and father, Hassabis grew up in North London within the Eighties. And thru the town haze, each once in a while, Hassabis might see one constellation—Orion, named for Greek mythology’s formidable hunter and for hundreds of years a information to sailors and farmers. Some 40 years later, it stays Hassabis’s favourite constellation, partly for its connection to the immortal: Even the traditional Egyptians honored Orion. 

“Initially, it’s a bit random, these patterns of stars all lined up, as we glance up from Earth,” Hassabis says. “And secondly, take into consideration Orion’s Belt: It’s three stars which might be simply randomly configured. However they imply one thing as a result of we’re utilizing our consciousness to interpret it.”

Hassabis and I are assembly not removed from the place he grew up—on the UCL Observatory, close to telescopes greater than a century previous and nonetheless raised to the sky. It’s a becoming place to speak about vastness, not simply of the celebrities however of ourselves. 

It’s additionally a becoming place to speak with somebody who’s well-known for devoting his personal consciousness to discovering that means in huge fields of information. Hassabis is among the most vital AI researchers and entrepreneurs of our time: He’s the cofounder of DeepMind, the pioneering AI lab that was acquired by Google in 2014. In 2016, DeepMind’s AlphaGo marked a seminal second in AI by defeating the world’s greatest participant in Go, one of many world’s most difficult two-player technique video games. Greater than a decade later, Hassabis leads Google’s core AI operations, serving to to steer the enormous at a time when it’s clawing its solution to the entrance of the aggressive pack on the power of its Gemini 3 mannequin.

However his most consequential work so far, maybe, is the improvement of AlphaFold 2—an AI system, unveiled by DeepMind in 2020, that might efficiently predict the three-dimensional constructions of proteins from their DNA sequences. AlphaFold 2 was a generational scientific achievement with implications for higher understanding and even curing ailments like Parkinson’s, muscular dystrophy, and sure cancers, all of which stem from misfolded or malfunctioning proteins. It gained Hassabis and DeepMind scientist John Jumper the 2024 Nobel Prize in Chemistry; that very same 12 months, Hassabis was knighted. 

To Sir Demis, it’s all linked. His early fascination with the skies has by traces to AI, discovering order and that means amid seeming randomness. 

“The night time sky is a thriller that’s staring us within the face on a regular basis,” he says. “It’s a continuing reminder of the larger questions. I feel that’s how I received into vastness…You’ve received to search out patterns in large quantities of information, or discover the proper transfer in large quantities of prospects.”

Hassabis, for the previous few years, has been devoting an vital share of his 100-hour workweek to one of many world’s biggest pattern-recognition issues: drug discovery. In 2021, with funding from Google dad or mum Alphabet, Hassabis began Isomorphic Labs, an AI drug-design firm that goals to create new, breakthrough medicines for among the most “undruggable” ailments—with the hyper-ambitious aim, because the startup’s splashy tagline places it, to “remedy all illness.”

Isomorphic has been quiet since launch and has but to maneuver a drug to the make-or-break medical trial part. However latest strikes recommend that milestone isn’t far off, and its backers argue that Isomorphic’s method will give it an edge as soon as it enters the fray. The startup lately opened its doorways to Fortune; I spent three days speaking with its executives and scientists about what’s arguably AI’s greatest alternative and problem. 

“A biotech startup may do one or two medication its whole company life,” says Hassabis. “However we’re making an attempt to construct a system, a course of, and all of the expertise to do possibly dozens of medicine every year. That appears loopy proper now, however I feel finally, over the following 10 to twenty years, we might get to discovering an answer to all illness…if we’ve a course of that may discover these needles in a haystack.”

Drug discovery is extra like discovering a needle in Iowa: It’s a means of testing doubtlessly therapeutic compounds towards the infinite variables of biology, characterised by continuous setbacks and an astronomical failure charge. 

Although it addressed solely a small a part of that course of, AlphaFold provided hope for a break from that actuality—among the first seismic proof that AI might take a brute-force, grind-it-out downside in medical science and compress a course of that when took years into minutes. After that breakthrough, Hassabis based Isomorphic with a easy thought: What if you happen to might flip AlphaFold right into a full-fledged drug-design engine? 

The ensuing spinoff goals to reach an area the place many have failed by specializing in construction: utilizing AI to generate detailed molecular-level predictions in regards to the interactions of medicine with their targets, thus stripping out a lot of the time-consuming trial and error that defines the pre-clinical-trial phases of drug discovery—and elevating the brash notion of “fixing” illness to the realm of the attainable. 

After its spinoff, Isomorphic initially raised cash from Alphabet, falling into the behemoth’s “Different Bets” bucket. In March 2025, the corporate raised a further $600 million, in a Sequence A led by Joshua Kushner’s Thrive Capital that included participation from Google Ventures, which was concerned from the inception. (Isomorphic declined to reveal the valuation.) The guess: that over time, we are going to design medication that remedy beforehand intractable ailments like most cancers and Alzheimer’s, with new techdriven processes so exact they appear virtually magical proper now—however that, finally, will grow to be normal. 

“Nobody would visualize designing an airplane at present by hand, nor would you wish to fly an airplane designed by hand,” says Thrive Capital accomplice Vince Hankes. “However all of our medication are designed like that. Sooner or later, they need to all be designed with sturdy software program and intelligence and simulation, similar to we design airplanes at present.” 

Isomorphic’s 300 or so workers are aiming to just do that, with Hassabis as, so to talk, their pilot.

Brutally lengthy odds

There are way more attainable chemical compounds than stars within the observable universe—about 10^60, or 10 to the sixtieth energy, based on the most recent analysis. That estimate covers solely small, drug-like molecules and will finally be low. Determining which of these combos may tame a tumor or harmful mutation is the duty Hassabis and his friends hope to unravel with AI. 

All through most of historical past, there have been only a few medication, and most of the ones that did exist had been found by chance. (Penicillin, found because of an unintended mould contamination, is probably the most well-known instance.) Within the Sixties, drug discovery picked up, as early most cancers and cardiovascular remedies emerged. However for a lot of the twentieth century, scientists scoured the chemical universe with a mixture of brute drive and slowly enhancing expertise. Many chemists spent their careers boiling sludge, operating lab checks, and ranging from scratch—and often failing. Even at present, based on broadly cited business figures, just one in 20 drug discovery chemists will efficiently carry a drug to market throughout their careers. 

“There are many completely different parameters you’re making an attempt to triangulate into one molecule that’s an ideal match for a particular downside,” explains Miles Congreve, chief scientific officer at Isomorphic. “You may discover that you just’ve received a terrific goal, it’s a potent compound, and it does very nicely. However there are different issues that aren’t proper—and also you go down a useless finish. It’s a bit like Whac-a-Mole.”

Congreve is anomalous amongst medicinal chemists: He has helped get three most cancers medication to market, together with Novartis and Astex Prescribed drugs’ ribociclib, which treats breast most cancers. Industrywide, even getting a drug to medical trials is usually thought-about a large win. However as he factors out, “Traditionally, there’s not less than a 90% failure charge” on such trials. “Your probabilities of discovering that excellent molecule are infinitesimally small,” agrees Fiona Marshall, president of biomedical analysis at Novartis.

These odds assist clarify simply how shocked a globeful of scientists was that AlphaFold 2 labored so nicely. That breakthrough, in flip, has helped Isomorphic appeal to expertise. Melissa Davis, director of computational biology, says she got here aboard exactly as a result of she was intrigued by constructing on AlphaFold. “Individuals would spend their complete profession making an attempt to crystallize one membrane protein,” notes Davis. “Abruptly, you didn’t need to spend 5 or 6 years making an attempt to get a construction for the protein anymore. Any scientist might generate one like that.”

Different prime employees have longer histories with Hassabis. Max Jaderberg—who served as Isomorphic’s chief AI officer for 4 years earlier than being named in November to succeed longtime Hassabis collaborator Colin Murdoch as president—spent seven years at DeepMind growing (amongst different issues) AlphaStar, the primary AI to greatest a human skilled on the online game StarCraft II. Jaderberg is outstanding amongst a cohort of DeepMinders who adopted Hassabis to Isomorphic (they make up about 11% of the corporate’s employees). 

“It’s at all times humbling to listen to you could have medicinal chemists who will do their complete profession with out making a single profitable drug,” says Jaderberg. “Distinction that with somebody like myself, who comes from the AI world the place you must smash one of the best on this planet each six months otherwise you’re useless.” He provides, “It’s humbling when rubber meets the street, with actual scientific processes and actual wetlab work.”

Getting the proper expertise is certainly one of Hassabis’s priorities, on condition that his crammed schedule limits his time at Isomorphic: He’s on the startup workplace at some point per week, often a Tuesday, when he meets with its government group and units priorities for the corporate’s technical route. 

Hassabis jokes that he loves managing “high-maintenance geniuses,” and that he’s on the lookout for these with a artistic streak. “Any skilled scientist will already be superb technically,” says Hassabis. “However then are you able to provide you with a artistic new thought, or ask the proper query? That’s truly more durable. Discovering the reply is definitely discovering the proper query.”

Construction first

What Isomorphic calls its structure-first method is, Jaderberg explains, a selection of generalization over specialization. The startup is specializing in an effort to map increasingly of the advanced organic constellations of the physique, the higher to foretell how any compound may have an effect on a spread of ailments and different organic processes. CTO Sergei Yakneen says it’s all about working towards a precision that when would have appeared unfathomable, like touchdown a rocket on the facet of the moon you’ll be able to’t see. 

Its core expertise is a drug-design engine constructed round numerous proprietary fashions. The engine incorporates an up to date protein-predicting mannequin, plus fashions for peptides, molecular glues, and antibodies. The info the engine is constructed on features a mixture of the worldwide Protein Knowledge Financial institution, the U.Ok. Biobank, commercially licensed sources, internally generated datasets, and information from companions. 

The duty is partially squeezing extra perception out of present information— one thing others have tried to do up to now, usually with out success, Yakneen acknowledges. “Then, lo and behold,” he provides, “with the proper expertise, you’re in a position to construct these mind-blowing techniques.” 

Isomorphic gained’t say what ailments it’s concentrating on within the quick time period—a secrecy that’s regular in pharma and barely odd in tech. The corporate factors to its partnerships with pharma giants Eli Lilly and Novartis as proof of its progress. (The Novartis partnership was expanded in 2025.)

In dialog, nevertheless, a number of executives say they’re centered on drugging the undruggable. This can be a broadly used phrase in drug improvement which means one thing comparatively particular: tackling protein mutations which might be notably prevalent in pancreatic, lung, and colorectal cancers, together with transcription components, that are widespread throughout varied most cancers varieties. All of those cancers have hitherto been proof against remedy, however they’re possible the sorts of codes Isomorphic is most dedicated to cracking.

Saving 5 years, or extra

Each drug discovery and AI economics are unforgiving. To get a brand new drug to market, you’ll possible spend greater than $2 billion and a decade or extra from discovery by medical trials—solely to face that 90% failure charge. In AI, in the meantime, you’re consistently operating up towards compute woes; there, Isomorphic’s Alphabet backing offers it some deep-pocketed help. 

Isomorphic additionally inhabits a viscerally aggressive market: The strain to be the primary startup to carry an AI-driven drug to market is intense. Rivals like Insilico and Recursion are making headway; at present Insilico has a number of medication in China-based medical trials. Isomorphic says it’s transferring towards trials, however declines to debate a timeline. One signal that day is nearer: the June 2025 hiring of chief medical officer Ben Wolf, a precision oncology professional. Wolf is recruiting a Boston-based group. “For this all to work,” he says, “I would like a brilliant medication, one thing with superior pharmaceutical properties that offers me the flexibility to check that it really works straightforwardly.” 

The startup for now could be staffed and oriented to focus totally on the drug-discovery course of, not medical trials or commercialization. On that entrance, Jaderberg is conscious of each the probabilities and the restrictions. “We’re at all times going to have, not less than over the midterm, components of biology which might be mysterious to humanity,” he says. The aim, he provides, is to “put scientific processes in place so it’s much less like magic, and extra such as you’re establishing mousetraps to isolate the consequences you’re making an attempt to drive.” 

Novartis’s Marshall sees a path for AI to hurry up discovery and trials by 50%. “I’d assume you possibly can get to 5 years’ common time,” says Marshall, including that many of the financial savings would come from an improved discovery course of. “I can’t see how we’re going to shave off way more than that, since you’ve nonetheless received human biology and security that wants doing” by medical trials. 

There’s a broad sense amongst medical scientists that AI drug discovery has over the previous decade promised greater than it may possibly ship—and Isomorphic is promising a complete lot. After I carry this as much as Hassabis, he outlines his philosophy: The concept of “fixing illness” is broader and extra sensible than eliminating sickness as soon as and for all. There’s a cause that he doesn’t say “remedy.” When you can’t promise nobody will ever get sick once more, he says, you’ll be able to develop a scientific, repeatable, and scalable course of—powered by superior AI and expertise platforms—for locating, designing, and optimizing medication or remedies as wants come up. 

“We’ll be build up our basic understanding of biology,” Hassabis says. “Hopefully we will provide you with one thing like a digital cell that’s predictive about what would occur if you happen to did sure interventions.” 

He reckons that might be attainable 10 years from now, which results in the following query: “How customized might it get?…You could possibly think about going right into a pharmacy and phenotyping your particular illness. So you recognize precisely what’s particular person to you”—a doubtlessly large breakthrough in illness remedy. 

Occupied with different universes, Hassabis believes, will help us start to grasp the organic one inside us. The phrase “isomorphic,” in any case, refers to 2 objects that seem completely different however are comparable in construction. 

After speaking with Hassabis, I walked over to the UCL Observatory’s Fry Telescope, which dates to 1862. Wanting by it, I noticed Saturn. It takes about 95 minutes for mild to journey between that planet and Earth, and it felt surreal to see one thing so distant, so clearly. 

“The universe is ready up by some means for science to work,” Hassabis had mentioned. “I really feel it virtually needs to be understood. In any other case, why would the scientific technique work so nicely and be so repeatable? Overlook AI…why ought to computer systems even work? They’re these bits of sand, steel, and electrons transferring round. After which one thing superb occurs.”

The drug discovery course of

Determine illness space: Researchers outline the illness to focus on and determine unmet medical wants the place new therapies might meaningfully enhance affected person outcomes.

Goal identification and validation: Scientists pinpoint the mechanism driving the illness and make sure that modifying this goal would profit sufferers. This stage prices about $1 million and may take from three months to 3 years. 

Assay improvement: Researchers design dependable laboratory checks that may precisely measure whether or not a compound impacts the goal within the meant approach.

Excessive-throughput screening to hit identification: Tens of millions of compounds are quickly examined utilizing automated techniques. This part often prices about $3 million and takes 12 to 18 months.

Hit: A “hit” is a compound that efficiently influences the goal in early testing. There could also be one promising hit—or a number of dozen.

Hit-to-lead optimization: Scientists refine the preliminary hits to enhance their effectiveness, selectivity, and drug-like properties. This part usually prices round $3 million and takes 12 to 18 months.

Lead optimization: Researchers additional refine the strongest candidate to maximise security, efficacy, and stability. This high-risk stage takes two to 4 years and prices between $5 million and $10 million.

Preclinical part: The drug candidate undergoes laboratory and animal testing to evaluate security, toxicity, and effectiveness earlier than human trials. This part lasts one to 3 years and prices $10 million to $20 million.

Investigational new drug (IND) software: Builders submit preclinical security information to the FDA and request permission to start testing the drug in people.

Part I-III medical trials: The drug is examined in progressively bigger teams of sufferers to judge security, dosage, and effectiveness. Scientific trials can take so long as a decade and value between $1 billion and $2 billion .

This text seems within the February/March 2026 challenge of Fortune with the headline “Google’s AI pioneer and his drug-design moonshot.”