Isomorphic Labs wants drug design on computers before testing in people

Max Jaderberg says a major problem in drug development is that toxicity or other failures may become visible only when a candidate reaches patients. His aim is to improve understanding before that stage, moving more discovery and design work onto computers.

Elad Gil introduces Jaderberg, president and a founding team member of Isomorphic Labs, an Alphabet-backed company working on AI-driven drug design. Jaderberg previously worked in machine learning at DeepMind.

At DeepMind, Jaderberg worked on generative models, language models, deep reinforcement learning and game-playing projects. By 2018–19, he says, he wanted to apply those methods to important problems beyond demonstrating what AI could do.

Early AlphaFold results gave the team confidence that deep learning could generalize in a meaningful area of biology. Jaderberg says the ambition to change how medicines are designed required a dedicated company bringing together multiple disciplines; Isomorphic Labs began around 2021.

Jaderberg traces his path through an engineering background and a computer vision PhD that began around the emergence of AlexNet. He worked on image recognition and text detection, started a computer-vision company, and joined DeepMind after approaching Demis Hassabis about investment.

Gil asks why drug approvals have remained roughly flat despite more scientific knowledge, modalities, time and spending. Jaderberg offers one explanation: medicine development has depended heavily on empirical trial and error, in which researchers must make and test molecules repeatedly to see what happens.

Historically, some medicines were discovered through observation before their mechanisms were understood. Jaderberg says biology and molecule design involve enormous, high-dimensional spaces; physical screening limits how much of that space researchers can search with finite time and money.

In his view, many comparatively accessible discoveries have been made, while more complex problems are harder to reach with the same process. Gil asks whether industry age, regulation and bureaucracy also contribute to slow progress, drawing a distinction between scientific and organizational explanations.

Jaderberg agrees that culture and the ability to keep innovating matter. He asks how research can gain economies of scale: what does work on one medicine teach a team that makes the next medicine faster? He describes many current programs as bespoke scientific projects with limited transferable learning.

His proposed alternative is a shared AI platform that absorbs learning across programs about disease biology and molecule design. He hopes accumulated data and model improvements can compound across medicines; this is the company’s strategy, not a demonstrated guarantee that every subsequent program will succeed.

Jaderberg says the intersection of frontier AI and drug discovery is relatively new, making ready-made experts hard to find. Isomorphic hires strong AI researchers and, separately, biologists and chemists, then works to build a culture where the disciplines learn from one another.

Gil describes a past genomics team with roughly five machine-learning people per domain scientist. Jaderberg says a similar ratio can fit some core AI modeling projects, provided domain experts understand the data and metrics well enough to prevent teams optimizing misleading measures.

Jaderberg says the company needs researchers for both simulation models and reasoning models adapted to the life-sciences domain. He then describes its north star: bring drug design and understanding of disease biology onto computers as much as possible.

Making molecules and testing them in physical assays takes time and money, he says. His goal is to make predictive models accurate enough to support more design iterations in silico, analogous to computer-aided design in other fields.

He uses AlphaFold’s protein-structure predictions as an example of a model that can accelerate part of the process. Isomorphic uses experimental accuracy as an internal target for its modeling work, he says; achieving it across the many distinct drug-discovery problems remains the goal.

Jaderberg expects better models to let researchers make and test fewer candidates in the physical world, increasing the chance that those selected will work. Gil asks how much of the long development process, especially the clinical portion, could realistically be represented on a computer.

Jaderberg outlines discovery, preclinical animal studies and clinical phases. In his account, clinical failures drive substantial cost: a candidate may reveal unexpected toxicity, bind an unintended target or affect its intended target in a harmful way once tested in patients.

The opportunity, he argues, is to understand disease biology, cells and potential effects before patients are exposed, rather than only predicting which molecule binds to which target. More predictive work is intended to inform selection; the conversation does not establish that it replaces clinical trials.

Asked about training data, Jaderberg says existing public and industry repositories can be valuable but may be biased products of earlier drug programs. Isomorphic also generates its own data, seeking distributions that better support generalization. He reports encouraging data scaling behavior without giving detailed results here.

Gil notes that results in mice often do not translate to humans and asks how the company obtains data about human side effects. Jaderberg says animal studies are an imperfect proxy and cites an approximate 30% figure for rat prediction of human clinical toxicity, while noting that estimates depend on the study.

Jaderberg proposes exploring combinations of laboratory measurements, AI models, organoids, organs-on-chips and synthetic human tissues to improve predictions of human toxicity. He emphasizes that clinical data are limited because only a finite number of trials have occurred. These approaches are presented as research opportunities, not established replacements for safety testing.

The drug-design models are built in-house, Jaderberg says, and the company sees value in owning proprietary data. It also works with contract research organizations and other companies because experimental platforms are diverse and difficult to build alone.

Gil asks about automated wet labs and feedback loops. Jaderberg says the company has worked with automated labs and supports reducing experimental turnaround, but some delays—chemical reactions or growing cells—have physical limits. He wants modeling to reduce how often researchers must wait for those cycles.

The desired division of labor is more prediction on a computer and physical experiments for validation. Jaderberg explicitly says real-world testing remains necessary to assess novel hypotheses and ultimately determine whether treatments work and are safe in people.

Isomorphic does not sell or license the model itself, Jaderberg says. It uses the engine for both internally funded drug programs and collaborations with pharmaceutical or biotechnology partners, which can bring disease expertise and longstanding difficult targets.

The company’s internal programs focus on oncology, immunology and inflammation. Jaderberg describes the design engine as intended to work across disease areas, target classes and treatment modalities, though the interview does not establish equal performance across all of them.

He wants to pursue mechanisms that might affect multiple downstream conditions. In cancer he points to a reported KRAS inhibitor example from Revolution Medicines as an illustration of a shared driver; he also notes that drugging such targets is difficult and tumors can evolve resistance.

Gil asks about combining multiple drugs, comparing cancer with HIV treatment. Jaderberg says individual medicines can have toxicity or interactions that make combinations difficult. Gil adds that testing combinations also raises the number of possible trials and the challenge of finding enough participants.

Jaderberg hopes more precise molecule design could eventually produce combinations that avoid harmful interactions and affect several targets across a biological pathway. He frames this as a future possibility, not evidence that such combinations have already been proven safe or effective.

Gil raises an argument about whether regulation places too much weight on safety relative to potential benefit. Jaderberg answers that patient safety must remain central, while the acceptable risk can differ depending on the condition and the options available.

He contrasts a patient with late-stage cancer and few options with someone receiving treatment for a lower-level chronic condition. He hopes that better-designed molecules will eventually improve safety and success rates, potentially changing regulatory discussions, but says that difference needs to be demonstrated over time.

Asked when Isomorphic will start its first clinical trial, Jaderberg declines to give a date. He says the company has medicines in development, is expanding its clinical team and wants to reach patients as quickly as possible; he does not report a completed or ongoing trial.

Asked for a prediction about treating a large share of cancers, Jaderberg does not give a date or percentage. He says he would be disappointed if, within five years, predictive understanding of disease and the ability to design more precise, less toxic molecules had not improved substantially. Drug development would still take years, he adds.

Gil turns to longevity, citing examples in worms, flies and mice where genetic interventions extend lifespan, and asks how to think about human life extension. Those animal examples are the basis of his question, not evidence that the same extension is achievable in humans.

Jaderberg says reducing major diseases could affect lifespan, while chronic age-related problems can also erode quality of life. Loss of bone and muscle strength, for example, can reduce mobility and have further effects even when a person lives many more years.

Gil distinguishes lifespan from healthspan, the period of life spent healthy, and describes animal aging models in which both increase. Jaderberg says researchers can pursue both relief of conditions that reduce healthspan and underlying processes of aging; he offers no estimate of human lifespan gains.