Can Drug Discovery Learn Across Projects? Max Jaderberg on Isomorphic Labs' AI Bet

1. Guest Background

Max Jaderberg approaches drug discovery from a background in building machine-learning systems. He is president and a founding-team member of Isomorphic Labs, whose work brings AI models together with biology and drug discovery. In this No Priors conversation, hosted by Sarah Guo and Elad Gil, he explains why the company wants more of the search for medicines to happen computationally, and what still requires evidence from the physical world.

Jaderberg describes starting in engineering, undertaking a computer-vision PhD and building a computer-vision-as-a-service company. Conversations with Demis eventually brought him to DeepMind. His work there included generative models, language models, deep reinforcement learning and difficult game-playing challenges. That experience shaped his conviction that apparently intractable problems could yield to systematic technical work, while raising a further question: where should those capabilities be applied?

He identifies AlphaFold as a turning point. To him, its results demonstrated that a deep-learning model could generalize within a meaningful biological domain. That supplied conviction for a dedicated company, which he places around 2021, with its own culture and specialists from different disciplines. The mission to solve disease explains the breadth of the discussion, from biopharma to longevity. It is an ambition framing the work, rather than a description of capabilities already established.

2. What the Episode Covers

The interview begins with an apparent productivity puzzle. The host describes an industry with expanding scientific knowledge and more kinds of therapeutic molecules, yet roughly flat new-drug approvals and increasing development burdens. He cites about fifteen years and $1.5 billion per drug. Those figures are his framing of the problem; the conversation does not establish a common methodology or show that they apply uniformly across programs.

Jaderberg's explanation centers on the limits of empirical search. Researchers formulate a biological hypothesis, design a molecule, make it, test it and revise. Cells are complex, high-dimensional systems, while small molecules, antibodies and peptides offer enormous combinatorial design spaces. A finite experimental budget can investigate only so much of that space. His argument is that accessible solutions are progressively found, leaving challenges that demand more than another round of the same search process.

Isomorphic's response is to move more disease understanding and molecular design onto computers. Jaderberg says internal model leaderboards include a target line for experimental accuracy. The intended benefit is to perform more exploration before making physical molecules, improving the choices that proceed to experiments. The target matters, but the episode does not supply model-by-model results demonstrating how much of the workflow has reached it.

The company develops its design engine and AI models internally. Data generation is more distributed: Jaderberg describes using its own work, contract research organizations and other companies with experimental platforms. This arrangement makes experimental capability part of the computational strategy. The models depend on measurements, even when their purpose is to reduce future rounds of measurement.

Commercially, he describes an internal engine used to make medicines, rather than a model, licensing or token business. Isomorphic pursues programs it owns and funds, alongside projects with pharmaceutical and biotechnology partners. Partners bring disease-specific expertise and difficult problems; Jaderberg's positive account of joint progress is not accompanied here by project results that independently establish its magnitude.

The internal pipeline focuses on oncology, immunology and inflammation. He also claims that the models apply across disease areas, target classes and modalities, although the interview does not demonstrate that breadth with separate evaluations. His project-selection logic favors shared disease-driving mechanisms, or root nodes, where a successful intervention might benefit multiple downstream indications. That connects the company's immediate drug programs to its larger platform thesis: can work on individual medicines build capabilities useful beyond each original project?

3. Core Views: Reasoning, Examples, and Limits

The strongest strategic claim is about accumulating knowledge. Jaderberg describes conventional drug programs as bespoke frontier-science projects: a team may learn a great deal about one target without becoming proportionately faster at the next. A shared AI platform could absorb experience across programs and improve its underlying understanding of disease and molecular design. The implied test is demanding. A growing portfolio alone does not demonstrate compounding R&D productivity; earlier work must improve decisions on later, meaningfully different problems.

His experimental-accuracy ambition also needs a precise unit of analysis. He uses AlphaFold as an example of obtaining structural information that might otherwise require lengthy laboratory work. That illustrates how prediction can replace a particular information-gathering step. But he also describes drug discovery as a collection of many difficult scientific and engineering problems. Success on structure prediction cannot establish equivalent performance on disease response, toxicity or clinical benefit. The unresolved issue is how far reliable prediction extends beyond each demonstrated task.

Laboratory automation clarifies why reducing experimental rounds could matter. Jaderberg welcomes better automated laboratories and says Isomorphic has worked with them for a long time. Yet chemical reactions still take time, and cells still need to grow into a testable state. Faster physical execution and fewer necessary physical iterations therefore offer complementary benefits. The second benefit depends on trustworthy predictions: omitting an experiment is useful only if the information needed for the next decision is adequately supported elsewhere.

Clinical failure raises the stakes beyond molecular binding. Jaderberg distinguishes a molecule that hits an unintended target from one whose intended action itself causes harm. Better selectivity could address the first problem while leaving the second unresolved. His reasoning therefore requires models of cellular behavior and disease mechanisms, alongside models of molecular interaction. A system can be right about what a molecule binds and still be wrong about whether manipulating that target will help patients.

That requirement changes the data question. Jaderberg argues that historical repositories often reflect the byproducts of previous drug-development activity. Their distributions may be useful for those projects without supporting a general solution. He emphasizes deliberately generating data suited to generalization and reports encouraging scaling behavior. However, he provides no dataset sizes, curves or quantified evaluations in this discussion. The supported insight concerns the importance of choosing what to measure; the claimed performance gains remain impossible to assess from the interview alone.

Human toxicity makes the evidence gap particularly visible. The host challenges the translation from animal results to people. Jaderberg cites roughly thirty percent predictiveness for rat-to-human clinical toxicity, while qualifying the figure by study. Without an identified study or metric, it should be read as his cited illustration, not a universal accuracy rate. He proposes combining in vitro endpoints, AI, organoids, organs on chips and synthetic human tissues. These could supply complementary signals closer to the human problem, but he also acknowledges limited clinical data. The proposal does not establish that those inputs already predict human toxicity adequately.

Combination therapy gives the design ambition a concrete application. The discussion uses RAS-related drug development to illustrate shared cancer mechanisms and notes that mutation can allow tumors to evade treatment. The transcribed drug name is uncertain, so this example cannot safely support a more specific account of a named therapy's results. Jaderberg argues that toxicity, liver effects and harmful interactions limit the combinations available today. More precise molecules might make simultaneous modulation of several targets feasible. The host adds a separate constraint: possible combinations multiply while patient recruitment remains finite. Improved molecular properties would not automatically solve the clinical evaluation problem.

Jaderberg's regulatory expectation is consequently conditional. He hopes that AI-designed molecules will eventually demonstrate statistically different success rates and safety profiles, influencing how regulation is approached. He presents that distinction as something to prove over time. The useful analytical question throughout the episode is therefore which uncertainty a technical advance has actually resolved, and which uncertainties remain for experiments and clinical studies.

4. Learning and Application

A practical way to assess this strategy is to separate the development stages before judging acceleration. Jaderberg distinguishes discovering a promising molecule, conducting preclinical studies and moving through phased clinical studies. He also emphasizes that failures contribute to average development costs. A reader evaluating a program should therefore ask what has improved: the time needed to choose a candidate, the evidence supporting that candidate, or outcomes in people. Progress in the first category cannot by itself establish progress in the last.

For a research team, the computational ambition can become a decision process with explicit responsibilities. Specify the property a model predicts, the choice that prediction will inform and the uncertainty that still requires an experiment. This can focus physical testing on the questions that matter most. Its boundary follows Jaderberg's own account: novel scientific hypotheses still require real-world validation, and treatments ultimately need evidence that they work safely in people. Confidence in a proxy measurement should not silently become confidence in the complete human response.

The organizational discussion offers another usable principle. Jaderberg describes recruiting strong AI specialists and strong biology and chemistry specialists, then building cross-disciplinary expertise through collaboration. A team can borrow that approach by bringing model development, data interpretation and medicine-making objectives into the same working conversation. It requires investment in mutual learning; assembling different credentials does not establish that the relevant knowledge is being shared.

His comments on staffing are narrower than a company-wide formula. He finds an approximately five-to-one ratio of modelers to domain expertise plausible for some core modeling projects. The more transferable point is the domain expert's responsibility: understanding data nuances and deciding whether a metric corresponds to progress toward a medicine. Teams applying this lesson could involve those experts when defining objectives and acceptance criteria, rather than only after a model has been built. The ratio should follow the task and its evidence needs.

Risk assessment also needs a defined patient context. Jaderberg contrasts late-stage cancer with few remaining options against low-level chronic inflammation. Applying that distinction means stating disease burden, available alternatives and the benefit being sought before discussing acceptable uncertainty. It does not establish a general safety exemption for AI-designed medicines; patient safety remains central to his argument.

Finally, track forecasts at the level actually offered. Jaderberg gives no first-trial date and does not commit to a timetable for treating eighty percent of cancers. His five-year expectation concerns predictive understanding and more precise, less toxic molecular design tools. Those milestones should be tracked separately from drug-development outcomes. The closing longevity discussion adds another distinction: lifespan and health span. Mobility, muscle and bone deterioration make functional years relevant alongside survival. Animal life-extension examples discussed by the host do not establish a human extension magnitude. A useful assessment would keep tool progress, treatment evidence and the kind of health benefit sought visible as separate questions.

Source

• Original episode: Powering AI-Driven Drug Discovery with President of Isomorphic Labs Max Jaderberg