To Solve All Diseases, Isomorphic Labs Needs a Half-Dozen AlphaFolds

Max Jaderberg's path through DeepMind ran through computer vision and generative models before reinforcement learning hooked him. At the time, he says, DeepMind was the place in the world to work on RL, and the question on everyone's mind was how to get an AI that could go off and do any task you wanted. The dominant paradigm was supervised learning, which trains a model by pairing each example with a known answer. That works when humans already know the solution. It fails on the problems that matter most.

Most of the world's hard problems, Jaderberg argues, are ones where humans do not know the answer. Drug design is a canonical case. Reinforcement learning is looser: you only need to say whether the model's output was good or not, maybe how good. That distinction looks small and opens an entire class of problems. DeepMind's early Atari work—Pong, Space Invaders—was the first big proof that the method could scale from toys toward something that looked like real problems in the world.

Host Stephanie Zhan notes that Sequoia was Atari's first investor. Jaderberg had no idea. He recalls researchers training agents on Pong and then sitting down to play a game themselves. The anecdote is a reminder of how far the same research thread has traveled: from arcade games in a London lab to a company trying to remake how medicines are designed.

Video games appealed to RL researchers because they are malleable, perfectly encapsulated worlds. You can manipulate them, test algorithms, and set up situations. After Atari, the instinct was always to find a more complex game. Jaderberg was obsessed with a different setup than the standard loop of train-on-this-game, then start over on the next. He wanted to train an agent once, lift it onto any new task, and have it perform without further training.

That demand is generalization over task space, which means you need lots of training tasks. In this regime, the training data is not images or text but tasks. The team hand-authored hundreds of mini-games, then procedurally generated worlds, maps, and objectives. They kept hitting a complexity ceiling: there is only so much complexity humans can design. Multiplayer games broke the ceiling. The moment another player is on the board, every opponent and every strategy fundamentally changes the game.

Why do professional chess players keep playing the same game for a lifetime? Because the opponents keep changing, Jaderberg says, so the game is never actually the same. Capture the Flag was DeepMind's first serious foray into stretching RL with multiplayer first-person games, and it reached human-level performance. StarCraft II and AlphaStar added another layer of complexity. Concepts the team discussed from 2015 through 2018—generalization, multi-agent dynamics, training on tasks—now sit at the center of large language models, at performance they could only dream about then.

What transferred was not a game engine but a set of building blocks. The same deep-learning primitives used in computer vision in 2012 showed up in early language models and in RL. Jaderberg describes the people who can combine those primitives as master chefs. Give them a hard problem that conference speakers call ten years away, plus enough compute, and you often find the team has already basically cracked it. That pattern, repeated enough times, is what convinced him the methods were ready for problems with real-world stakes.

Jaderberg has worked with Demis Hassabis for more than a decade, since Google acquired both DeepMind and Jaderberg's company Vision Factory around 2014. Hassabis, he says, has a "rollout mind"—the chess player's habit of playing out which present moves lead to the largest future ambition. The habit showed up at a toast in an early DeepMind lobby, celebrating the lab's first Nature paper: the Atari DQN work, essentially the first machine-learning paper in the journal. Hassabis told the room it would be the first of dozens, and that they would win Nobel Prizes as a result. Ten years later, Hassabis shared the 2024 Nobel Prize in Chemistry for AlphaFold.

Isomorphic Labs, spun out of DeepMind, inherited that scale of ambition. The goal is not a drug for one indication. It is solving all diseases: using AI to understand biology, craft chemistry that modulates it, and open new pathways at the root of disease. From day one the company was organized around a general drug-design engine—something that can be applied again and again across disease areas, targets, and even modalities.

That choice inverts what Jaderberg calls the first generation of AI in drug design: local models. A team fits a small network to a scrap of data about one target or one chemical class, then uses it to guess the next design round. Isomorphic wanted models that generalize across chemistry and across target space, the way AlphaFold 3 can be pointed at any protein in the proteome and any small molecule you can think of, without fine-tuning. Every internal research program is aimed at that kind of generality. To assemble the breakthrough engine, Jaderberg says, they need something like half a dozen AlphaFolds. Structure prediction is only part of the story.

A pill has to be absorbed, reach the right cell, enter it, and survive the liver. Each of those is an AlphaFold-scale problem: experimental-level accuracy on a core concept of biology or chemistry. AlphaFold 2 cracked protein structure; AlphaFold 3 extended that to proteins with small molecules, DNA, and RNA. The remaining breakthroughs are about the rest of what a designer has to hold in their head—properties, function, and the journey a molecule takes through the body.

Even a suite of holy-grail predictive models—experimental-level, or better, forecasts of a molecule's properties and of real experiments—would not solve drug design. The number of possible drug-like molecules is often cited as 10^60. Cut that by twenty orders of magnitude and you still have 10^40. Screen a billion candidates a year, 10^9, and 10^31 molecules remain on the table. Prediction answers whether a given molecule is any good. It does not tell you which molecule to look at next.

That is why Isomorphic is building generative models and agents on top of its predictors. Jaderberg compares molecular design to Go, not chess: chess can be searched exhaustively, Go cannot, and AlphaGo had to combine generation with search. Agents that use generative models, search, and predictive scoring are how you navigate 10^40 or 10^60 space without exhausting it. The surprising fact, he notes, is that humans found drugs in that space without AI at all.

His reading is redundancy. For a given target and product profile, many designs should work. Humans have opened only a tiny corner. Generative models and agents could uncover the rest—the needles in a haystack that might be life-changing for millions of people. In the language of games that still shapes how he thinks, you need a world model of biochemistry and then a player that knows how to explore it.

AlphaFold 2 predicted single-protein structures. AlphaFold-Multimer added protein complexes. Therapeutics still required a further hop. Small-molecule drugs—caffeine, paracetamol, things you can often take as a pill—work by attaching to proteins and disrupting or enhancing the molecular machines those proteins form. A designer's job is to invent a molecule that fits a protein in a particular way. Seeing the three-dimensional interactions used to mean crystallizing the complex in a lab, which can take six months, years, or prove impossible.

AlphaFold 3 predicts how proteins interact with small molecules, DNA, RNA, and antibody–antigen interfaces, in one model. That structural view opens target classes that were hard to drug, including transcription factors that sit on DNA and read genes. At Isomorphic, designers sit with a browser-based interface, change a design, and immediately see how the physical interaction changes. Ligands and small molecules are the same problem under two names; biologics such as antibodies are another mode of the same model.

Getting there required new tokenization and a new architecture. Proteins were a known input. DNA and RNA could be tokenized as bases. For small molecules the team tried many schemes and found atomic-resolution tokenization worked best. AlphaFold 2's framework could not predict mixed molecular systems, so a diffusion model took over, generating the 3D coordinates of every atom, conditioned on those tokens. On the internal leaderboard, small-molecule–protein interaction accuracy jumped by a generation, unlocking the rest of the project.

Demis Hassabis has said biology is not data-constrained. Jaderberg's version is more precise: every machine-learning field feels a data constraint, but here it is not a real bottleneck. Progress is possible with data that already exists and data the team can generate; nobody has to wait fifty years. There are modeling problems where data has been sitting around for years and models can now leap past anything people have seen. That does not mean data is unimportant. The historical record in biology was not created for machine learning, and going out to generate data in order to train a model is a different activity.

Jaderberg has been a synthetic-data believer since his PhD, when a few thousand images against Google's millions forced him to generate data himself. Chemistry has theory to lean on: quantum chemistry and quantum mechanics, approximated into scalable molecular-dynamics simulations, plus generative models scored to raise information content. The true gap is in vivo data—measurements on a mouse or a rat—which you cannot fabricate at scale. He is watching organoids-on-a-chip as a way to measure animal-like readouts without the animal. Isomorphic has no labs of its own; it partners with other companies and generates a lot of proprietary data.

Drug development's bottlenecks do not stop at structure. Clinical trials—how to test molecules in people, quickly and safely—will have to change as predictive models of toxicity get better. A pill goes everywhere, and toxicity often comes from hitting proteins other than the intended target. If a general model can predict interaction with one target, the same model should predict interaction with the rest of the proteome. Internally, the first programs are in immunology and oncology. Partnerships with Eli Lilly and Novartis brought targets the field had worked on for more than a decade; the models helped uncover new chemical matter, and the Novartis collaboration later expanded.

World experts in drug design who are also world experts in deep learning barely exist, because the fields have not coexisted long enough. Isomorphic sits the two groups side by side and asks them to speak each other's language with curiosity. Some 60 to 80 percent of the machine-learning team has no prior chemistry or biology beyond school, which Jaderberg calls an asset: naive, "stupid" questions force first-principles thinking and punch through dogma. In November 2024 the company launched the AlphaFold Server, with code and model weights for non-commercial academic use. Researchers are already using it in creative ways far beyond drug design.

Jaderberg remembers GPT-3 as the crossing where generated text started to look as if a human had written it: the model had begun to recreate the distribution it was trained on. In biology, the analogous moment is a generative model that emits molecular designs that exist in the physical world, or could exist—stable, manufacturable, real. Unlike language, chemistry and biology are domains humans struggle to understand even when they are looking at the truth. So the moment will feel less like GPT-3 and more like AlphaGo's Move 37 against Lee Sedol.

Move 37 was uninterpretable. It looked like a mistake. No one had played it in thousands of years of Go, and it turned out to be the critical move of the match. Isomorphic is already seeing a version of that internally: a generative model proposes a design, a human drug designer prefers their own, and the wet-lab test shows the model was right and the human was wrong. Jaderberg expects a great deal of that behavior as models are applied outside native human understanding.

Asked when the first AI-generated drug will enter the clinic, he turns the question toward process. As AI-designed molecules reach clinical phases, the industry will need new ways to engage regulators and to fold predictive models of on-target action and off-target toxicity into trial design. Trials themselves may eventually be rethought as design gets faster, more targeted, and better informed. Traditional pharma, he says, will not sit on the other side of an "AI versus pharma" line. There is no world, five years out, in which you design a drug without AI. It will be like trying to do science without mathematics.