Powering AI-Driven Drug Discovery with President of Isomorphic Labs Max Jaderberg
Summary
- Drug development’s rising costs reflect the limits of empirical trial and error. Gil describes roughly flat annual approvals despite 10–20 times more publications, with development reaching about $1.5 billion and 15 years. Jaderberg argues the “low-hanging fruit” has been found; the opportunity is identifying problems before expensive clinical failures.
- Isomorphic wants drug-design iteration to move onto computers at experimental accuracy. AlphaFold illustrates the potential: predicting structures can replace work taking months or years. Jaderberg’s ambition is to “go beyond the real world,” concentrating physical experiments on validation while acknowledging that human safety still requires testing.
- The platform’s potential advantage is cumulative learning from deliberately generated data. Jaderberg argues foundation models can “generalize at scale,” allowing one program’s learning to improve the next. Historical pharmaceutical datasets can be “very, very biased”; Isomorphic generates data suited to broader prediction and reports encouraging scaling behavior.
- Isomorphic keeps its AI engine internal and develops owned and partnered medicines. “We do not sell this model,” Jaderberg says. Internal priorities are oncology, immunology and inflammation; pharmaceutical partners contribute expertise and difficult problems they may have pursued for 10–20 years.
- More precise molecules could enable safer, more effective cancer-drug combinations. Gil highlights tumor mutations and the precedent of HIV cocktails; Jaderberg identifies toxicity, liver effects and drug interactions as constraints. If those properties can be engineered precisely, combinations could modulate “a whole vector of targets.”
- The five-year benchmark is a major advance in predictive biology and molecular precision. Jaderberg would be “very disappointed” without that progress, while discovery still takes years. He gives neither a first clinical-trial date nor a timetable for curing 80% of cancers; Isomorphic is expanding its clinical-development team.
- The longevity discussion centers on healthier years and aging’s root causes. Gil says statistics “seem to indicate” that eliminating cancer and heart disease adds five to seven years, contrasting that with threefold worm lifespans after gene interventions. Jaderberg sees opportunities in treating age-related ailments and underlying deterioration, without forecasting a human lifespan multiplier.
Deep dive
1. Drug discovery needs learning that carries between projects
Gil’s puzzle: publications increased 10–20 times and therapeutic modalities multiplied, yet annual drug approvals remained roughly constant. He puts development at about $1.5 billion and 15 years, versus several hundred million dollars and eight or nine years previously. The expanding scientific toolkit has not delivered a corresponding increase in output.
Jaderberg’s diagnosis is that drugmaking remains “a very empirical process”: hypothesize, make a molecule, test, repeat. Biology and molecular design are enormous, multidimensional search spaces; finite budgets limit the complexity trial and error can reach. He thinks the “low-hanging fruit” accessible through that process has already been found.
Gil pushes beyond science to bureaucracy, regulation and industry aging, noting the scarcity of newer $50 billion biotechs outside Moderna. Jaderberg agrees culture matters, then asks what transfers between programs. Treating “each project as a stand-alone” limits cumulative progress; foundation models offer the ability to “generalize at scale.”
AlphaFold supplied the founding conviction: evidence that deep learning could generalize across a significant biological space. After working on seemingly impossible gaming challenges around 2018–19, Jaderberg wanted to apply those capabilities to consequential problems. Isomorphic began around 2021 with a separate company and culture designed to combine disciplines and move quickly.
2. Experimental accuracy could move iteration onto computers
Jaderberg’s organizing question is “How can we go beyond the real world?” Automated laboratories help, but chemistry and cell growth retain physical waiting times: reactions must happen, cells must reach the required state. Computer-aided design transformed other engineering disciplines; he wants drug discovery to undergo that shift.
Internal leaderboards carry a line marking “experimental accuracy.” AlphaFold illustrates the payoff: obtaining structures through X-ray crystallography can take months or years. If predictive models reach experimental precision across discovery problems, more design can happen virtually before molecules are synthesized. Physical and ultimately human validation remain necessary.
The data strategy follows the generalization goal. Public, pharmaceutical and corporate repositories often reflect passive collection during earlier projects and can be “very, very biased.” Isomorphic generates its own data and reports “some really amazing data scaling laws.” It combines internal capabilities, CROs and specialist companies to supply those datasets.
The talent model combines strong AI researchers, potentially without life-sciences backgrounds, with biology and chemistry specialists. Gil proposes roughly five machine-learning experts per domain scientist; Jaderberg says that resembles core modeling teams. Domain expertise keeps researchers from “chasing false metrics,” while working together develops interdisciplinary judgment.
3. Human toxicity is the expensive prediction problem
Gil’s economic pushback is specific: if discovery and early experiments consume roughly $50 million and two to three years, while clinical work consumes about $1.45 billion and 12 years, how much can computation really change? Jaderberg’s answer is that failures drive average costs upward—and the consequential failures emerge in patients.
Binding the intended target is insufficient: off-target activity, unexpected toxicity or harmful consequences of the intended mechanism can still defeat a drug. “Oh no, there’s a toxicity we didn’t account for.” Gil presses the translation problem: cancer has been cured in mice hundreds of times without producing effective human treatments.
Depending on the study, Jaderberg puts rats’ predictive performance for human clinical toxicity at about 30%. His proposed alternative combines measurable in vitro signals, AI models, organoids and organs on chips to predict human outcomes more directly. Clinical data remain limited; this is an opportunity to build better predictors, rather than a claim that the problem is solved.
4. The business develops medicines and pursues difficult cancer biology
“We do not sell this model,” Jaderberg says: Isomorphic neither licenses its engine nor sells access. It applies the technology to fully controlled, self-funded programs and pharmaceutical partnerships. Partners bring disease expertise and problems pursued for 10–20 years. Internal priorities are oncology, immunology and inflammation, while he describes the engine as unrestricted by disease, target type or therapeutic modality.
Gil introduces Revolution Medicines’ KRAS inhibitor as an example, saying it dramatically improved the lives of patients with late-stage pancreatic cancer and received a standing ovation at a major cancer conference “earlier this year.” He frames it as an example of tackling root causes. Jaderberg acknowledges the point.
Gil’s pushback: tumors contain a distribution of mutations, so suppressing one vulnerability can leave others to expand. HIV treatment illustrates the case for cocktails. He also flags the clinical combinatorics: limited trial participants make testing numerous combinations difficult.
Jaderberg locates another constraint in imperfect molecules: toxicity, liver effects and harmful interactions make combinations unsafe. If drugs can be designed more precisely around those properties, combinations could influence “a whole vector of targets” across complex pathways. That possibility depends on better molecular engineering.
5. Safety and predictive tools advance on different timelines
Gil questions whether safety constraints suppress warranted risk-taking, citing wartime penicillin commercialization and COVID-vaccine speed. Jaderberg keeps patient safety central but distinguishes late-stage cancer from chronic low-grade inflammation: acceptable risk depends on patient needs. He hopes better mechanistic design eventually produces “statistically significant differences” in safety and success, creating evidence that changes regulation.
Asked for the first clinical-trial date, Jaderberg declines to specify one and says the clinical-development team is expanding. Asked when 80% of cancers become curable, he instead sets a tools benchmark: “I would be very disappointed” without a huge advance in predictive disease biology within five years. More precise, less toxic molecules are part of that ambition; drug discovery still takes years.
Gil says statistics “seem to indicate” eliminating cancer and heart disease adds five to seven years. He contrasts that with 20–30% lifespan gains from gene interventions in animal models and threefold gains in C. elegans, including longer healthy adulthood. Jaderberg emphasizes the final 10–20 years of declining health and sees opportunities both to treat ailments and address aging’s root causes. He offers no forecast of threefold or fivefold human lifespans.