Powering AI-Driven Drug Discovery with President of Isomorphic Labs Max Jaderberg｜No Priors — BigGo Finance

Max Jaderberg, president and founding-team member of Isomorphic Labs, argues that drug discovery has stalled because it remains an empirical trial-and-error process, and that moving disease biology and molecule design onto computers is the way to break the logjam. The conversation spends most of its time on why pharma productivity has flatlined despite exploding knowledge, what “experimental accuracy” in simulation models would mean for cutting the 15-year, $1.5 billion development cycle, and how Isomorphic is organized and staffed to pursue that. Longevity and healthspan come up only briefly, near the end.

Jaderberg came out of DeepMind, where he worked on AlphaGo and AlphaStar after selling a computer-vision company to the lab, and he frames Isomorphic as the application of that general-purpose deep learning to biology, with AlphaFold as the proof of concept. Host Elad Gil, who holds a biology PhD and worked on aging pathways, presses on the parts of drug development that resist simulation — animal models that poorly predict human toxicity, clinical failure, and the data problem underneath it all. The reader should come away with Isomorphic’s operating model (own the models, own some programs, partner on others), its stated focus areas, and the specific numbers Jaderberg uses to describe the industry’s failure modes.

Pharma’s productivity has flatlined because discovery is still empirical

Jaderberg’s core diagnosis is that medicine-making has never escaped trial and error. Early drugs were stumbled upon — he notes aspirin and paracetamol were used for a long time before anyone understood how they worked — and the industry still runs on hypotheses tested iteratively in the real world. The problem is dimensional: the cell is “a big complex chaotic system,” and molecules (small molecules, antibodies, peptides) are combinatorially vast design spaces. The only tools available have been screening and trial-and-error, which caps the complexity of any solution reachable within a finite budget or timeline.

Gil lays out the paradox with numbers: new drug approvals per year are roughly flat even as the scientific literature grew 10–20x, the number of drug modalities expanded (antibodies, small molecules, proteins, peptides), and cost and time both worsened — from a few hundred million dollars and 8–9 years per drug to about $1.5 billion and 15 years. Jaderberg’s read is that the low-hanging fruit findable by empirical processes is being exhausted, and each remaining problem demands more complexity than the process can reach.

He also points to a structural cause on the business side: drug programs are run as bespoke frontier science projects, so learnings from one program transfer poorly to the next. Without economies of scale in R&D, the compounding that builds large companies never kicks in. Gil adds the industry-age argument — older industries slow down through bureaucracy and regulatory capture — and notes it has been roughly 40 years since a new tens-of-billions-dollar biotech was built, citing Regeneron, Vertex, and Amgen as the 1980s-era cohort, with Moderna’s rise treated as almost a COVID accident.

“Where things really fail is in the clinic because we never could see the problems until we get to the point where we put these drugs in patients and we say, ‘Oh no, like there’s toxicity that we didn’t account for.'”

Most clinical failures are visibility failures, and simulation is the fix

Jaderberg breaks the 15-year, $1.5 billion cycle into stages and argues the cost is driven by failure rates, not by any single expensive step. Discovery plus preclinical work (animal studies for safety and efficacy) is a small slice; the bulk is Phase 1 through Phase 3. Failures cluster in the clinic for three recurring reasons: unanticipated toxicity, off-target effects where the molecule hits something else with negative consequences, or on-target effects where hitting the intended target itself harms the patient.

The opportunity, in his framing, is to understand those problems before patients are exposed — modeling not just “what molecule binds to what” but the surrounding disease biology and how modulating a target actually plays out in cells. Gil’s challenge is the data: cancer has been “cured” in mice hundreds of times without translating to people, and the data that matters most comes from human trials. Jaderberg concedes the point sharply on animal models — a rat is roughly 30% predictive of human clinical toxicity, he says, and no AI scientist could ship a model at that accuracy. He proposes replacing the animal proxy with a combination of in vitro endpoints, AI predictive models, and newer experimental platforms such as organoids, organs-on-a-chip, and synthetic human tissues, then modeling human clinical toxicity directly — while acknowledging the data there is thin because only a finite number of trials have run.

The north star is “experimental accuracy” — get out of the real world

Isomorphic’s stated goal is to move drug discovery and design work onto a computer, and its internal metric is experimental accuracy: a predictive model good enough that you no longer need the physical experiment. AlphaFold is the template — structure prediction that once took months or years of X-ray crystallography, now done in silico, saving that time outright. Jaderberg says every leaderboard inside the company carries a dotted line for experimental accuracy, and the aim is that when Isomorphic does go to the lab, it is validating rather than learning.

He is explicit that automation of wet labs is complementary but not the answer. Robotic and automated labs have improved a lot and Isomorphic has worked with them for a long time, but physics sets a floor: a reaction takes as long as it takes, and culturing a cell means waiting days for the right state. Gil raises the closed-loop argument — fast feedback is what makes AI models improve — and Jaderberg’s response is to push as much of the loop as possible into silicon so real-world time is spent validating novel hypotheses, not iterating on them.

On data, Jaderberg describes a “ridiculously comprehensive” strategy built on a key lesson from the past five years: generate it yourself. Public and pharma repositories are byproducts of years of passive collection and are heavily biased, which hurts if the goal is generalization. Isomorphic’s own data generation, he says, has produced strong data-scaling laws.

Isomorphic owns the engine and runs two kinds of programs

The commercial model is vertically integrated and deliberately closed. Isomorphic builds the AI drug-design engine in house and does not sell the model, license it, or sell tokens — it is used internally to make medicines. Those medicines come through two channels: wholly owned and self-funded internal programs, and partnered programs with pharma and biotech companies that bring deep expertise in specific disease areas, sometimes after 10–20 years without progress on a hard problem.

Internally the focus is oncology, immunology, and inflammation. The models themselves are described as general — disease-agnostic, target-class-agnostic, and modality-agnostic — which is what lets partnered work range across disease areas. Within oncology, Jaderberg favors “root node” problems whose solution could benefit many patients across downstream indications. His example is KRAS: Gil notes RAS has appeared in cancer literature for 20–30 years, and Jaderberg cites Revolution Medicines’ KRAS inhibitor, which drew a standing ovation at a major cancer conference for results in late-stage pancreatic cancer. He flags the difficulty specific to cancer — tumors mutate around solutions — which is why he is interested in combinations.

Combinations are blocked by imprecise molecules, not by biology

Gil draws the HIV analogy: AIDS was suppressed with drug cocktails, and tumors contain a spectrum of mutations that likely need simultaneous suppression or the disease evolves around treatment. Jaderberg agrees and names the binding constraint — drug-drug interactions. Individual molecules often work but carry liver effects or toxicity, so stacking several of them harms patients, which caps how complex a combination can be tested and used. The fix he proposes is better engineering: molecules precise and clean enough that combinations become safe, enabling regimens that modulate a whole vector of targets across complex pathways rather than one at a time. He frames this as an expansion of the applicable surface of medicine, not just an efficiency gain.

Safety trade-offs should flex with the patient, not the regulator

Gil relays Jan Janssen’s argument that drug development overweights safety and that regulators curtail progress, citing the wartime rediscovery and commercialization of penicillin in about nine months once regulatory obstacles were removed, and the COVID vaccine as a modern parallel. Jaderberg does not endorse loosening safety standards. He says patient safety must stay central, but that risk appetite legitimately varies by indication — late-stage cancer patients with few options accept risks that a chronic low-grade inflammation patient would not. He says Isomorphic talks regularly to the FDA and regulators worldwide and finds shared intent to get patients the best drugs. His bet is that statistically better success rates and safety profiles from designed molecules will eventually change how regulation is calibrated.

Hiring is the binding constraint on a field that barely exists

Jaderberg says you cannot hire world experts in AI for drug discovery because the field is too new — and the few who exist use different techniques. Isomorphic therefore hires the best core AI and deep learning people, some with no life-science background but who want their work to have human impact, alongside core biologists and chemists with little AI exposure, and builds a culture where the two learn from each other until cross-disciplinary experts emerge. Gil’s own experience at Color was roughly five ML people to one traditional research scientist; Jaderberg says that one-in-five ratio resonates with Isomorphic’s teams. The domain experts’ job is to keep modelers from chasing smoke metrics and to ensure results translate into real medicines. Active hiring is concentrated on core AI researchers, both for simulation models like AlphaFold++ and for adapting frontier and open-source reasoning models to the biology space.

Longevity is a secondary frame; healthspan is the operative one

The longevity discussion is short. Gil notes that removing cancer and heart disease might add only five to seven years of average life, while knocking out certain genes in C. elegans, Drosophila, or mice extends life 20–30% (threefold in the worms), and that in aging models healthspan extends alongside lifespan rather than trailing it. Jaderberg’s response stays close to disease: major killers have concrete impact on how long people live, and beyond them the chronic degradations of aging — bone and muscle loss, a fall that ends mobility, the mental-health knock-on — erode the last 10–20 years of life. He endorses attacking both the afflictions and the root causes of cellular and organ degradation, without committing to any life-extension multiple.

What to watch

The episode’s unresolved tension is between Jaderberg’s “get out of the real world” thesis and the parts of drug development that only humans can adjudicate. His own 30%-predictive rat statistic is the strongest argument in the transcript for why simulation must improve — and also a reminder that the validation step never disappears. Two concrete markers will show whether the thesis is working: Isomorphic’s first clinical trial, which Jaderberg declines to date beyond saying the company is building out clinical development and pushing hard, and his stated five-year test — that predictive understanding of disease biology and the ability to design precise, low-toxicity molecules will have step-changed enough to start cracking large portions of disease. If neither has moved by then, the empirical-process critique will apply to Isomorphic too.