Isomorphic Labs Bets on Simulated Experiments to Design Better Drugs

Bloomberg's Shirin Ghaffary introduces Max Jaderberg, president of Isomorphic Labs and a member of its founding team. The company began in 2021 with the ambition to rethink drug discovery using AI and better understand the mechanisms of life.

Jaderberg came to the work from engineering, computer vision and general AI rather than biology or drug development. During his Oxford PhD and an early company, he worked on the basic question of whether AI models could be made useful at all.

He met Demis Hassabis, sold his company to DeepMind and joined it. Over the following years, he saw AI move from a promising idea to systems that could master games such as Go and StarCraft, and then make progress on protein folding.

The protein-folding results prompted him to reconsider the purpose of building general technology. Seeing a path to an important real-world problem, he decided to focus on human health and the discovery of new therapies rather than wait for artificial general intelligence.

Although drug development was new to him, Jaderberg says his father's medical and drug-development background was an influence he appreciated more as he became immersed in the field. He has stayed focused on it since.

Ghaffary asks how Isomorphic Labs intends to make money, given its partnerships with companies including Eli Lilly, Novartis and Johnson & Johnson. Jaderberg says its goal is to make medicines for patients: it uses its models internally rather than selling tokens or API access.

The company has two paths. In its own programs it develops ideas, conducts science, designs molecules, seeks patents and aims to prove candidates in clinical trials. It also works with drug companies that bring years of expertise and a biological rationale for a possible medicine.

The economics differ. A pharmaceutical partner can pay an upfront fee, development milestones and later royalties. For an internal program, Isomorphic funds years of research itself; if evidence becomes strong enough, it can consider licensing, partnering or commercializing a medicine itself.

Asked how close the company is to commercialization, Jaderberg describes it as early in its life. Several programs are in preclinical development: researchers design molecules and test them in wet labs, cells and animal models before deciding whether there is enough evidence to proceed to trials in patients.

He says the AI drug-design engine is intended to work across diseases and types of treatment, including small-molecule pills, antibodies and peptides. Developing a particular treatment still requires specialized clinical and biological expertise for the disease involved.

For programs it fully funds and runs itself, the company concentrates on oncology, immunology and inflammation. Jaderberg says those areas still contain significant unmet patient needs; partnerships allow the company to apply its technology in other fields with specialists there.

Ghaffary raises a common objection: even if AI accelerates research and suggests drug candidates, trials, laboratory work and regulatory approvals remain bottlenecks. Jaderberg says the time taken by safety regulation has a good reason: patients need protection.

He argues that conventional discovery has relied heavily on trial and error, such as testing huge numbers of molecules against cells. In his account, this limits the complexity of medicines researchers can design, and many candidates that enter clinical trials still fail because they prove toxic or do not work as intended.

Jaderberg believes better models of biology and disease, together with AI-designed molecules, could eventually improve the success rate of trials. He says any resulting change in the process should be gradual, with patient safety foremost; candidates still have to be tested carefully in people.

Asked about competition from frontier AI companies, he says he has felt pressure from the start to deliver useful medicines quickly. He expects AI-designed drugs to become widespread, but does not want progress to take decades while patients are waiting.

Reasoning models can help integrate data and generate scientific hypotheses, he says. The difficulty is choosing which ideas deserve scarce laboratory time, since making molecules and carrying out experiments in cells take real effort.

Isomorphic also works on simulation or world models, including models of the kind used to understand protein folding. Jaderberg says approximating parts of the physical world on a computer can help researchers run virtual experiments and choose stronger candidates before committing to real experiments.

He expects the combination of reasoning and simulation to improve the odds when researchers move into laboratory work, preclinical studies and clinical trials. That is his prediction for the approach, which still depends on testing its results in the real world.

Ghaffary asks about his long working relationship with Demis Hassabis. Jaderberg says their shared values have sustained the partnership and that Hassabis, already enthusiastic about Isomorphic, has increased the attention he gives the company.

On the next year of AI in science, Jaderberg says mathematical and physics results have been striking even though he had expected progress. He expects advances in biology too, but not as immediately: a mathematical proof can often be checked on a computer, while a claim about a cell must be tested in a cell. Simulations can help, but laboratory time remains necessary.