Drug development's constraint isn't strict approval but invisible preclinical toxicity
Isomorphic Labs's thesis: drug cost inflation stems from clinical failures caused by preclinical blindness to toxicity—rat testing predicts human toxicity with only ~30% accuracy.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Drug development's constraint isn't strict approval but invisible preclinical toxicity
Over decades, scientific papers increased tenfold to twentyfold and new drug modalities—antibodies, small molecules, proteins, peptides—proliferated, yet annual approvals remained flat. Development costs ballooned from millions to $1.5 billion, timelines stretched from 8–9 years to 15, and every metric worsened. Max attributes this to an industry still trapped in experience-driven trial-and-error: hypothesize a molecule, then iterate in the real world. But molecular and biological design spaces are high-dimensional and combinatorially explosive; empirical search only captures low-hanging fruit.
— Max JaderbergThe goal is experimental-grade accuracy, not faster experiments
Isomorphic's north star isn't running wet experiments faster but moving drug design entirely to the computer. Max uses AlphaFold as the exemplar: structure prediction once required months or years of X-ray crystallography; if the model reaches experimental accuracy, real-world validation becomes unnecessary. Internal leaderboards mark an ‘experimental accuracy’ baseline; once crossed, wet experiments shift from ‘learning and iteration’ to ‘verification’, drastically reducing cycles in the physical world.
— Max JaderbergClinical trials are where drug failure and costs concentrate
Unpacking the 15-year, $1.5 billion timeline: small-molecule discovery and in-vitro work occupy only 2–3 years and ~$50 million; the rest is clinical. Most failure modes—unexpected toxicity, off-target effects, or mechanisms that hit the target yet harm the body—only emerge when the drug reaches patients; animal models and in-vitro tests miss them entirely. Real gains in failure prevention come from understanding these failure modes before patient exposure.
— Max JaderbergGenerate proprietary data instead of relying on public databases
Public and industry-wide data are mostly passive byproducts of historical collection, biased, and unfit for training general models demanding generalization. Isomorphic generates its own required data distributions instead of relying on existing large databases. Max noted this approach revealed ‘surprisingly impressive data scaling laws’ in model training.
— Max JaderbergRat toxicity testing predicts human outcomes with only about 30% accuracy
Regulators mandate animal proof of safety before human use, yet toxicity frequently emerges once drugs enter patients. Max argues directly: rats aren't humans; this ‘gold standard’ predicts human clinical toxicity with about 30% accuracy—and he half-jokes that if his team's AI hit 30%, he'd ask engineers to ‘do better’. He advocates combining in-vitro experiments, AI models, and new tissue-mimicking techniques—organoids and organs-on-chip—to model human clinical toxicity directly, bypassing the five-to-six-decade-old, unreliable animal proxy.
— Max JaderbergCombination therapy fails at safety engineering, not biology
HIV and cancer treatments use cocktail approaches—tumors contain mutation spectra; single drugs prompt resistance; hitting multiple targets simultaneously prevents escape. But real-world combination therapy is rare. Max explains: existing molecules are often intrinsically toxic or damage the liver; multiple drugs generate harmful interactions. The constraint isn't biological understanding but molecular design precision—molecules lack the refinement and safety. Better-engineered, cleaner molecules unlock multi-target therapy against complex pathways.
— Max JaderbergAcceptable risk varies by disease; safety standards shouldn't be one-size-fits-all
When asked whether safety-focused regulation slows drug development, Max grants patient safety must remain central but notes risk tolerance is dynamic: COVID's pervasiveness and severity shifted social acceptance of expedited review; late-stage cancer with no alternatives and chronic low-risk disease demand vastly different risk tolerance. True progress means faster access to better treatments, not uniform safety thresholds across all disease contexts.
— Max JaderbergExtending lifespan isn't enough—drug development must also improve later-life quality
Beating cancer and heart disease still leaves aging itself—accumulated chronic problems: bone and muscle loss, post-fall immobility, compromised wellbeing and quality of life. Decades past 80 aren't inherently well-lived. Max emphasizes the industry must work both directions: treat discrete fatal diseases to extend lifespan, and tackle aging's root mechanisms to extend healthspan.
— Max JaderbergIn their own words · checked verbatim
The large part of the industry makes medicines is this very empirically driven process.
Max Jaderberg0:00
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.
Max Jaderberg0:00
a rat isn't actually a human. It's not actually very predictive. I think, you know, choose your study, but it's about 30% predictive of human clinical tox, like in a rat.
Max Jaderberg20:53
if one of my, like AI scientists or engineers came up to me and said, hey, we've got this amazing model, it's 30% predictive. You're like, okay, maybe do a bit better than that.
Max Jaderberg20:53
I'd be really disappointed if we fast forward, you know, even five years time, if we don't have the base tools, like the predictive understanding of disease biology has not stepped change in like a massive way.
Max Jaderberg34:26
even if people live to a lovely long age of 80 or 90, like, they, maybe the last, like, 10 years of their life or even 20 years of their life isn't necessarily as great as it could be.
Max Jaderberg35:42
Figures
| Average drug development cost (today) | $1.5 billion | 6:09 |
| Average development timeline | 8–9 years → 15 years | 6:09 |
| Rat toxicity test accuracy for human clinical toxicity | ~30% | 20:53 |
Glossary
- AlphaFold
- DeepMind's protein structure prediction model—to Isomorphic, the first proof that biology can be generalized by deep learning.
- Organoids/organs-on-chip
- Synthetic human tissue systems that mimic organ microenvironments, used to replace animal toxicity testing.
- CRO (contract research organization)
- Third-party firms handling outsourced experimental work for pharmaceutical companies, with inconsistent quality.
How to listen
Founders and investors focused on AI drug discovery, biotech investment, or drug development processes, seeking to understand which stages of development AI can actually transform.
The opening 1–5 minutes covering Max's background and how he joined DeepMind has low information density and can be skipped.