Max Jaderberg

1. Key Themes

Drug discovery is stuck in an empirical, bespoke paradigm, and that is why productivity has been flat despite exploding knowledge

Elad Gil framed the puzzle: approvals per year are roughly flat while papers, modalities, and costs have all gone the wrong way ("It went from $100, a few hundred million dollars per drug to, um, 1.5 billion. It went from eight or nine years to 15 years." 00:05:45). Max Jaderberg's diagnosis is that the industry's core method hasn't changed: "the how the large part of the industry makes medicines, is this very empirically driven process. Empirical, as in you need to... have a hypothesis on, you know, how a disease might work... but you have to make it and test it in the real world and do this very iteratively." 00:07:18 The consequence is a ceiling on complexity: "the low-hanging fruit of medicines that can be found with our empirical process of designing medicines are being found... but then it's getting harder and harder to sort of ratchet up the complexity." 00:08:28

The goal is to move drug discovery "out of the real world" and onto a computer, to experimental accuracy

Jaderberg's stated North Star: "how can we bring all of that drug discovery work, all of that drug design work, understanding of, you know, disease biology, onto a computer? So, how can we get out of the real world?" 00:00:00 The bottleneck he identifies is physical: "the big time bottleneck, the big cost bottleneck, is synthesizing real atoms, and then testing them out in real assays." 00:15:10 He compares it to other engineering fields that moved to computer-aided design, and says each internal leaderboard has a target line: "internally, we have... a dotted line on all of our leader boards that is experimental accuracy... At that point, you don't need to go into the real world." 00:16:02 The real world remains necessary for validation, but not for learning: "it looks like we go into the real world to validate, not necessarily to learn and iterate." 00:25:50

Clinical failure, not discovery, is where the economics break, and the opportunity is predicting what only patients currently reveal

Gil noted the bulk of cost and time sits in clinical development, not the roughly "$50 million" early phase. Jaderberg agreed: "Why do things cost so much? Well, it's cuz so many things fail... where things really fail is in the clinic... we never could see the problems until we get to the point where... we put these drugs in patients, and we say, 'Oh, oh, no. Like, there's toxicity that we didn't account for.'" 00:18:00 The thesis: "the really big opportunity is to have much more of that understanding before you go into patients." 00:00:00 That means modeling cell biology, off-target effects, and on-target-but-harmful effects, "not just like what molecule will bind to what." 00:19:25

Compounding learning across programs is the real economy of scale an AI platform can offer

Jaderberg asked what traditional pharma learns from one drug that speeds up the next: "a lot of the techniques and the way we go after medicines... is really treating each one as like a segmented... bespoke frontier science project." 00:10:10 The AI alternative: "if you... have this platform underneath going after these medicines that allow you to like... suck up and understand the learnings and help impact your core foundational understanding of disease biology and how to design good molecules... then you can hope to actually get compounding improvements the more medicines you go after." 00:10:52 This is the generalization-with-scale dynamic of foundation models applied to biology.

Proprietary, purpose-built data beats legacy repositories

Existing datasets are biased by-products of past drug programs. Jaderberg: "oftentimes these repositories of data, you know, whether they're public or in pharma... have been the byproduct of many years of passive... data collection... So, they can be very, very biased, which often doesn't really help you if you're trying to make a general solution to drug discovery." 00:19:34 Isomorphic's response: "we go out and generate this ourselves, and we've seen some really, really amazing uh data scaling laws as a result." 00:20:14 Models are built in-house; experimental data generation is partnered broadly: "we work with the whole world, whether that's... stuff we do ourselves, contract research organizations, CROs, um, other innovative companies, to actually help us create the data that is... powering our models." 00:23:14

Business model: use the engine internally, don't sell the model

"We don't sell the model, we like we don't license out the model, we don't sell tokens in that way. We use it internally to make medicines." 00:26:27 Two tracks: wholly owned programs and partnerships where "they come to us with like, 'Hey, these are the hardest problems, maybe problems that they've been working on for sometimes 10, 20 years, not making much progress. And then together, using our drug design engine, we can make a huge amount of progress.'" 00:26:39 Internal focus is "oncology, immunology, and inflammation," while the engine is "disease area agnostic... target class agnostic... modality agnostic." 00:27:30

Hiring recipe: cross-disciplinary, AI-first, with roughly one domain expert per five AI people

Because "the world experts at AI for drug discovery... basically don't exist," Isomorphic hires "literally just the best people in core AI and deep learning, who maybe don't even have a background in life sciences," plus biologists and chemists new to AI, then builds culture so "everyone talks and learns from each other." 00:12:03 On team ratios, Jaderberg said Gil's one-in-five ratio "resonate[s] quite a lot for what our teams look like," with domain experts there to "make sure that we're not... chasing smoke metrics." 00:13:31

Better molecule engineering unlocks combination therapies

Cocktails are limited not just by trial combinatorics but by molecule quality: "it's actually because we're not great at engineering exquisite molecules and medicines that means that we can't combine them together in a safe way for patients." 00:30:20 The aspiration: "much more accurately engineer these molecules to be safe, to not have these effects on the liver, to not interact with each other in bad ways, then you can feasibly come up with... quite complex combinations that can modulate not just a single target at a time, but a whole vector of targets." 00:30:44

2. Contrarian Perspectives

Animal toxicology, the regulatory gold standard for 50-70 years, is a weak proxy that should be replaced

Jaderberg: "the gold standard that... you need to get through for regulators is you need to prove that your molecules aren't toxic on rats and other animals... a rat isn't actually a human. It's not actually very predictive. I think... it's about 30% predictive of human clinical tox." 00:21:04 His stance is to "forget animals. They've been a proxy for the last, you know, 50, 70 years. But we can maybe do better" by combining in vitro endpoints, AI models, and "organoids, organs-on-a-chip, like synthetic human tissues." 00:21:46

Better automation of wet labs is complementary, but not the endgame; physics limits it

While others build robotic closed-loop labs, Jaderberg argues the endgame is fewer real-world turns: "you're going to hit the finite laws of physics at some point. Like, how long does it take to synthesize a molecule?... How long does it take to culture a cell? At some point in time, you just have to wait those days." 00:24:45 His conclusion: "we should also be driving really hard at bringing that in silico on a computer so that we can reduce the amount of real-world time." 00:25:46

Safety regulation isn't the main villain; engineering better molecules will change the regulatory conversation

Gil raised Paul Janssen's view that overemphasis on safety constrains progress. Jaderberg declined to take that bait: "I think it's like centrally important to just always have patient safety in mind," with risk appetite differing by indication (late-stage cancer versus chronic inflammation). 00:32:30 Instead, he expects data to move regulators: "I hope to prove out over time that this is very statistically... different. And that will change the way that we think, I think, about regulation and getting drugs to patients." 00:33:38

Don't license AI models to pharma; keep them as an internal engine

In an era where "sell tokens" is the default AI business model, Isomorphic explicitly refuses: "we don't license out the model, we don't sell tokens in that way. We use it internally to make medicines." 00:26:27 This keeps the compounding learning and proprietary data inside the company.

Cocktails are underused because of molecule imperfection, not clinical-trial economics

Gil suggested trial combinatorics as a reason cancer drugs aren't combined; Jaderberg reframed it as an upstream chemistry problem (drug-drug interactions, liver effects). 00:30:20 That implies precision molecule engineering is the unlock for combination approaches, a different lever than trial design.

3. Companies Identified

Isomorphic Labs

Alphabet-backed AI drug discovery company, with Max Jaderberg as president, founded on DeepMind's AlphaFold lineage. Mentioned as the subject of the episode: it builds an in-house drug design engine, runs internal oncology/immunology/inflammation programs, and partners with pharma. Jaderberg on its origin: "AlphaFold results... this was like the first proof of concept that you could get a deep learning model that generalizes in... a meaningful space in biology." 00:02:32 On its goal: "going after solving all disease." 00:02:32 On its posture: "we've been building up our clinical development team, so we're getting ready for this." 00:34:42

Google DeepMind

AI research lab that acquired Jaderberg's computer-vision startup and incubated AlphaGo, AlphaStar and AlphaFold. Mentioned as the research origin of Isomorphic and the place where the proof-of-concept emerged. Jaderberg: "I was at DeepMind for a long time beforehand... AlphaGo, AlphaStar, um, and it got to the point where... we were going after these like 'impossible' problems." 00:01:56

AlphaFold (DeepMind / Isomorphic)

Deep learning model for protein structure prediction. Mentioned as the foundational proof of concept and the template for "experimental accuracy" simulation models. Jaderberg: "AlphaFold looks at structure prediction... Normally, that takes months or years to do X-ray crystallography... If you can do that with a simulation model like AlphaFold, then you don't have to go into the real world and do that." 00:16:02 The team is working on "AlphaFold plus plus plus" 00:14:19.

Revolution Medicines

Oncology company developing RAS-pathway inhibitors. Mentioned as a validation of "root node" targeting in cancer. Jaderberg: "daraxonrasib, like KRAS inhibitor, standing ovation at a big cancer conference like drastically improving... the lives of... late stage pancreatic cancer patients from Revolution Medicines." 00:28:30

Anthropic

AI lab. Gil noted it "recently announced that they're doing wet lab in San Francisco," reading it as a likely robotic closed-loop system. 00:23:56 Mentioned as a comparison point for the automated-experimentation approach.

Regeneron, Vertex, Amgen

Large-cap biotechs. Mentioned by Gil as the last generation of tens-of-billions-dollar biotech builds ("almost 40 years since we've had a new tens of billions of dollar biotech built"). 00:09:14 Jaderberg added "Regeneron, Vertex" 00:09:22; Gil added "Amgen was that era too." 00:09:26

Moderna

Gil called it "almost an accident of COVID" as the sole exception to the lack of new $50B+ biopharma companies. 00:09:14

Janssen Pharmaceuticals (J&J)

Pharma company founded by Paul Janssen, acquired by J&J. Mentioned by Gil as one of the most productive drug developers historically, and as the source of the argument that overemphasis on safety constrains progress. 00:31:19

Color

Genomics company Gil founded and led as CEO early on, applying ML and data science to genomics pipelines. Mentioned as the source of the one-research-scientist-to-five-ML-people team ratio. 00:13:23

4. People Identified

President and founding team member of Isomorphic Labs; Oxford computer vision PhD; sold a computer vision company to DeepMind (~2014); worked on AlphaGo, AlphaStar, generative and language models, and deep RL. Mentioned as the guest. His ImageNet-winning work: "We ended up having the uh ImageNet winning network of 2014, best object recognition, text detection." 00:04:26

Demis Hassabis

DeepMind co-founder; helped found and run Isomorphic. Mentioned as the person who recruited Jaderberg and co-drove the AlphaFold-to-Isomorphic conviction. Jaderberg: "I ended up just speaking to Demis Hassabis, um, asking for angel investment, and he was like, 'Well, why don't you just come to DeepMind?'" 00:04:26

Yann LeCun

Pioneer of convolutional neural networks. Mentioned as an early influence on Jaderberg's translation of convnets into computer vision research. 00:04:23

Elad Gil

Host; investor; started Color; holds a biology PhD with work on aging pathways. Mentioned as the interviewer who contributed domain framing (flat approvals, mouse-to-human translation failure, longevity genetics). Example: "we cured cancer in mice hundreds of times, and those molecules wouldn't translate into people." 00:20:30

Paul Janssen

Founder of Janssen Pharmaceuticals; drug development pioneer. Mentioned by Gil for the view that "the main constraint on drug development is an overemphasis on safety," citing the wartime rediscovery and commercialization of penicillin "in a nine-month period." 00:31:19

Winston Churchill

Cited by Gil as the leader who removed regulatory obstacles to get penicillin commercialized during World War II. 00:31:19

Alexander Fleming

Discoverer of penicillin in the 1920s; mentioned by Gil as the discovery that was "put on a shelf and forgotten." 00:31:19

5. Operating Insights

Put an "experimental accuracy" dotted line on every model leaderboard

Jaderberg turns a vague ambition into an operating metric: each AI capability is benchmarked against what the wet-lab experiment itself delivers, so the team knows exactly when a model can replace a lab step. "We have a dotted line on all of our leader boards that is experimental accuracy... At that point, you don't need to go into the real world." 00:16:02 The tactic generalizes to any team replacing a costly physical or human process: define the substitution threshold up front.

Pair every AI modeling pod with a domain expert to guard against "smoke metrics"

Rather than siloing, Isomorphic embeds domain experts in the modeling projects so metrics map to real medicine-making. Jaderberg: "you just need some key domain experts, people who really understand the application, the nuances of the data, like what do the metrics mean, um, make sure that we're not, you know, chasing smoke metrics." 00:13:31

Interrogate the predictive value of the "gold standard" before optimizing against it

Jaderberg applies the same bar to legacy endpoints that he'd apply to an internal model: "if one of my... 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, we can maybe do a bit better than that.'" 00:21:04 Operating lesson: audit whether the benchmark your industry treats as sacred actually predicts outcomes.

Commission your own data distribution when passive data is biased

Rather than relying on inherited data, design the data distribution for generalization. "If you really care about generalization out of these models, there's maybe a different distribution of data you need. So, we go out and generate this ourselves." 00:20:14

6. Overlooked Insights

Combination therapy is gated by molecule quality, a thesis that quietly expands the addressable surface of precision AI design

In a quick exchange, Jaderberg recasts why we don't give multi-drug cocktails for cancer, despite HIV's precedent: individual molecules carry liver toxicity and drug-drug interactions. "It's actually because we're not great at engineering exquisite molecules and medicines that means that we can't combine them together in a safe way for patients. And so, we're maybe limiting like the the surface of applicability." 00:30:20 If true, a precision-engineering platform doesn't just make single drugs better; it unlocks an entire category (rational combinations that outrun tumor mutation) that is currently structurally inaccessible, with the end vision of modulating "a whole vector of targets that span these like very complex biological pathways." 00:31:08

Human clinical toxicity prediction is a data-starved but tractable modeling target that sits at the intersection of three emerging technologies

Almost in passing, Jaderberg sketches a stack: in vitro endpoints, AI predictive models, and new organoid/organ-on-a-chip tech, aimed directly at modeling human clinical toxicity: "there's probably an opportunity to bring all of this stuff together and then directly try and model human clinical tox. Now, there's not much data on that side because... there's only a finite number of clinical trials that have been run, but I think we have some really good methods now." 00:22:26 Since clinical-stage failure is where the industry's cost explodes, any validated replacement for the roughly 30%-predictive rat model would be among the highest-leverage assets in drug development, and organoid/organ-on-chip companies are the natural supply chain for it.