Isomorphic Labs’ Max Jaderberg on narrowing drug designs with AI

Max Jaderberg, president of Isomorphic Labs, says his team uses AI models to narrow large numbers of molecular designs before making and testing a much smaller set in the laboratory. He speaks with Marina Mogilko about drug discovery, scientific judgment and how to tell whether an agent is useful.

Jaderberg’s background is in computer vision and general AI research at DeepMind. He says early AlphaFold results convinced the team that models could generalize in biology and chemistry, giving them a reason to pursue drug design with AI.

AlphaFold is a proof of possibility, not a complete medicine-making system, he says. A drug must reach an appropriate target, work as intended, reach the right place in the body and be safe. He says drug development is long and failure is common.

Isomorphic Labs thinks a collection of advances will be necessary, rather than a single model. Jaderberg describes modeling disease biology, molecular design and toxicity as distinct capabilities that would need to work together in a drug-design engine.

The work has two sides: understand how healthy and diseased biological systems behave, then design an intervention. Candidate medicines can include small molecules, antibodies or peptides, each intended to change a biological system in a useful way.

The company did not want to develop its models in isolation for a decade before testing commercial relevance. Jaderberg says it started drug-design work early, including partnerships with pharmaceutical companies, alongside internally funded programs whose intellectual property it owns.

Whether a technology is ready depends on the domain. Jaderberg says drug discovery is already an empirical process with imperfect predictions, so a model can be useful before it is perfect if it meaningfully improves the choice of experiments.

His team compares model benchmarks with live discovery programs. When a score improves but real scientific outcomes do not, he says the benchmark may be measuring the wrong proxy and needs to be redesigned around the practical problem.

The mission of solving disease is too large for one person to hold in mind, Jaderberg says. He breaks it into a tree of smaller missions and individual goals so people can connect daily work with the long-term ambition.

For one modeling task, he describes setting an 80% target when the existing result was around 10%. Even if distant, a numerical goal can show incremental progress and focus a team. He also acknowledges that such milestones can be artificial proxies.

Jaderberg sees opportunities beyond biology wherever progress depends on expensive physical experiments. Reasoning models may suggest hypotheses, while specialized models could simulate part of the physical world and help decide which experiments to perform.

He says available data is a key constraint: it must contain information relevant to the prediction, rather than being merely plentiful. Noisy or inadequate datasets may limit a model, but new ways to generate or collect data can create an opportunity.

Isomorphic Labs brings together AI researchers and engineers with chemists, biologists and clinicians. Jaderberg says he hires strong specialists without requiring all of them to have biology experience, provided they are curious and willing to learn across disciplines.

In interviews, he asks candidates about entering an unfamiliar field or encountering a technical failure: how did it feel, how did they approach learning, and what did they do next? He is looking for evidence of how they handle uncomfortable learning situations.

Generative models and design agents can propose many possible molecular designs. Other models act as a virtual laboratory, predicting responses and narrowing the list. The team then makes selected molecules and tests them in the real world, where experiments take time and money.

Jaderberg says that instead of experimentally screening hundreds of thousands of designs, the team can make 20 or 50 and obtain a high proportion of experimental hits. He does not give an exact hit rate or say that these laboratory hits have become approved medicines.

He sees researchers’ roles changing: an AI researcher can specify a space of model changes and an objective, while agents propose and run individual experiments. A chemist can similarly define the design space and desired properties for a molecular-design agent.

Jaderberg says he uses several models, including Gemini, to analyze and visualize company data. He gives this as a productivity example separate from the specialized models used for drug design.

Asked about drug-design agents that run for days, he says the prompt, surrounding software and definition of success all matter. Drug design involves several properties that can conflict, so the objective is not simply to maximize one number.

He warns that an agent can look impressive without saving time or producing better work. Teams should define what good performance means, evaluate parts of the system and measure whether it is actually better or faster than the alternative. In biology, he says, verification is often softer than in problems fully checkable on a computer.

Jaderberg says talking with and managing people is a large part of his job that he does not think should currently be automated. He also sees human attention, time and physical resources as constraints even when models can generate many hypotheses.

On regulation, he says evidence and patient safety are essential before testing a candidate in people. Early clinical trials begin cautiously to assess safety and tolerability; the need for careful testing remains even when earlier predictions improve.

Asked when AI-designed drugs will reach pharmacies, Jaderberg says he does not know. He says some medicines with machine-learning involvement are entering clinical studies and describes Isomorphic Labs’ own oncology, immunology and inflammation programs as in development.

He expects disease understanding and medicine design to improve over the next decades and sees potential for more personalized treatments. He does not provide a firm date for individual therapies or claim that a particular clinical outcome is assured.

His advice to people entering AI for science is to learn the science as well as the tools. Compute, human attention and laboratory time remain finite, so researchers still have to choose which questions deserve those resources.

He calls the ability to select useful research questions and promising approaches “research taste.” In his view it develops through hands-on experience, domain knowledge and intuition; ideas from other disciplines can also help a researcher see a problem differently.