By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Powering AI-Driven Drug Discovery with President of Isomorphic Labs Max Jaderberg

Drug discovery and development has long relied on expensive, time-consuming empirical trial and error.

Max JaderbergguestElad Gilhost

Nov 12, 202639mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Why drug discovery fails late: toxicity and unseen biology in patients
Max Jaderberg opens by describing drug discovery as largely empirical and explains that major failures often only appear once drugs reach human trials. He frames the biggest opportunity as gaining a deeper predictive understanding of biology before exposing patients to experimental medicines.
- •Drug discovery is still heavily trial-and-error despite scientific progress
- •Many programs fail in the clinic because issues like toxicity aren’t visible earlier
- •Off-target effects and unexpected biology emerge only when drugs are tested in people
- •The core opportunity is shifting more understanding earlier in the process
Isomorphic Labs’ origin story: from DeepMind breakthroughs to a new company
Elad Gil introduces Max and the company’s ambition, then Max recounts how work at DeepMind—especially early signs of AlphaFold’s success—helped convince the team that deep learning could generalize in biology. He explains why the mission required a standalone company with its own culture and interdisciplinary team.
- •Max’s DeepMind experience centered on “impossible problems” like game-playing systems
- •AlphaFold was viewed as proof that deep learning could generalize in meaningful biology spaces
- •The team wanted to redirect frontier AI toward health and disease
- •Isomorphic Labs was formed around 2021 to move fast with a dedicated culture and team
Max’s path into deep learning: computer vision, a startup, and joining DeepMind
Max traces his background from an engineering mindset into a computer vision PhD around the emergence of AlexNet. He describes building an ImageNet-winning network and a computer-vision-as-a-service startup that ultimately led to a conversation with Demis Hassabis and a move to DeepMind.
- •He began a PhD in computer vision just before AlexNet shifted the field
- •Worked on convnets and methods leading to strong results like an ImageNet-winning network
- •Built a company and sought funding, which led to meeting Demis Hassabis
- •Joined DeepMind after being invited to work toward AGI
Why drug approvals stay flat: empirical processes, complexity, and poor knowledge transfer
Elad lays out the paradox of flat drug approvals amid exploding scientific knowledge and rising cost/time. Max argues the core challenge is that biology and molecule design are high-dimensional and the industry relies on slow trial-and-error, with limited transfer of learning from one program to the next.
- •Drug discovery is slow because it requires making and testing molecules in the real world iteratively
- •Biology and molecule design are high-dimensional, combinatorially large problems
- •The “low-hanging fruit” gets found first; harder problems remain out of reach of empirical methods
- •Many programs are treated as bespoke projects with limited compounding learning across efforts
- •A platform approach could enable transferable learning and economies of scale in R&D
Building teams at the AI–biology intersection: hiring and cross-disciplinary culture
Max describes the hiring reality: few people are already experts in both frontier AI and drug discovery. Isomorphic hires excellent AI researchers and strong biology/chemistry domain experts, then relies on culture and collaboration to create cross-disciplinary expertise internally.
- •The combined field is young, so fully formed “AI for drug discovery” experts are rare
- •Isomorphic recruits top AI/deep learning talent even without life sciences backgrounds
- •They also hire top biology and chemistry experts who may not have AI experience
- •Internal culture is designed to help disciplines learn from each other
- •For core modeling work, a small number of domain experts guide larger modeling teams (roughly one in five)
Moving drug design onto computers: simulation models aimed at experimental accuracy
Max explains Isomorphic’s “North Star” as bringing drug discovery work onto computers to avoid the bottlenecks of synthesizing and testing real atoms. He emphasizes building predictive ‘world models’ with a benchmark of experimental accuracy, so real-world work becomes validation rather than iterative learning.
- •Core goal is to shift drug discovery/design from real-world iteration to computation
- •Real-world bottlenecks include synthesis and assays/experiments
- •Computer-aided design transformed other engineering fields; Isomorphic wants the analog in biology
- •Models are evaluated against a target of experimental accuracy
- •If models are predictive enough, fewer real-world cycles are needed and success probability improves
Why clinical trials drive cost: failure modes and the need to model more than binding
The conversation shifts to where time and cost accumulate, especially in clinical phases where failures are common. Max argues failures often happen because unknown toxicity, off-target effects, or incorrect biological assumptions only surface in humans, motivating broader modeling of disease biology beyond single-target binding.
- •Drug development costs are amplified by high failure rates, especially in the clinic
- •Failures include unexpected toxicity, off-target interactions, and harmful effects from the intended mechanism
- •Better predictive understanding of cells and disease biology could surface issues earlier
- •The goal is modeling multiple interacting components, not just a molecule binding to a target
- •Earlier insight can reduce wasted clinical attempts and overall timelines
Data strategy: generating less-biased datasets for generalization
Elad asks how Isomorphic gets the data required to model complex biology. Max describes a broad data strategy that includes generating proprietary data, noting that many existing repositories are biased because they are byproducts of historical drug-making rather than designed for generalizable modeling.
- •Isomorphic pursues a comprehensive data strategy spanning many sources
- •Public and industry repositories can help, but may be biased due to passive collection
- •Bias can limit generalization, so different data distributions may be needed
- •The company generates data itself to better support generalizable models
- •Max notes seeing strong “data scaling laws” from internally generated datasets
Rethinking safety endpoints: beyond animals toward human-relevant signals
Max and Elad discuss the limitations of animal models for predicting human outcomes. Max highlights that animal toxicity tests can be weak predictors of human toxicity and suggests combining lab measurements, AI models, and newer experimental systems like organoids and organs-on-chips to better approximate human biology.
- •Animal models can fail to predict human toxicity even when regulators require them
- •Max cites low predictiveness of rat toxicity for human clinical toxicity
- •He argues for rethinking endpoints and aiming to model human clinical toxicity more directly
- •Potential ingredients include in vitro endpoints plus AI predictive models
- •Newer experimental systems discussed include organoids, organs-on-chips, and synthetic human tissues
Internal vs external dependencies: in-house models, plus CROs and platform collaborators
Elad probes how much Isomorphic builds internally versus relying on outside groups. Max says the AI drug design engine and models are built in-house and that proprietary data is valuable, but experimental biology is long-tailed and complex—so they also work with CROs and other innovative companies to create needed data.
- •AI models and the drug design engine are developed entirely in-house
- •Proprietary data generation is viewed as a major opportunity to own critical inputs
- •Experimental biology platforms are complex and specialized, encouraging external collaboration
- •They work with CROs as well as other companies developing relevant experimental platforms
- •External work complements internal capabilities to power the models with data
Automated labs vs in-silico ambition: speeding loops but escaping physical bottlenecks
Elad raises the value of closed-loop wet labs and rapid iteration, noting long experiment cycle times in biology. Max agrees automation is improving and useful, but stresses physical constraints remain—reactions and cell culture still take time—reinforcing Isomorphic’s drive to do more in-silico work and use wet labs mainly for validation.
- •Automation and robotic labs can reduce iteration time and should be pursued
- •Biology still faces hard physical limits (reaction times, cell culture durations)
- •Max reiterates the goal of moving work onto computers to reduce reliance on real-world iteration
- •He differentiates validating results in the real world from learning/iterating in the real world
- •The aim is to minimize experimental turns by doing more predictive design up front
Business model and execution: building medicines (not selling the model) via internal programs and partnerships
Max explains that Isomorphic does not sell or license its model as a standalone product. Instead, it uses the engine internally to create medicines through both wholly owned programs and collaborations with pharma/biotech teams that bring deep disease-area expertise and hard, long-standing problems.
- •The company uses its AI engine internally rather than selling/licensing it
- •Two development paths: fully owned internal pipeline and partnered programs
- •Partners may bring problems they’ve pursued for many years with limited progress
- •Collaborations combine partner disease expertise with Isomorphic’s drug design engine
- •The focus is on making medicines rather than distributing the model itself
Where they start in health: oncology, immunology, inflammation—and the promise of general models
Max shares Isomorphic’s internal focus areas while emphasizing the generality of the underlying models. He notes the models are disease-area, target-class, and modality agnostic, enabling work across different therapeutic approaches even as the company prioritizes certain areas first.
- •Internal focus areas include oncology, immunology, and inflammation
- •He emphasizes the models are general rather than limited to one disease or target type
- •The system is described as modality agnostic, supporting multiple therapeutic modalities
- •They aim to tackle “root node” problems that could affect many downstream indications
- •The initial focus is strategic rather than a limitation of the platform
Complex disease requires complex therapies: multidrug combinations and toxicity constraints
Elad and Max discuss why combination therapies are less common in some contexts despite their success in areas like HIV. Max argues that limitations in molecular precision and safety—especially toxicity and drug-drug interactions—make combinations risky, and that improved design could enable safer, more complex multi-target approaches.
- •Combination therapy can be important when diseases evolve or contain heterogeneous mutations
- •A barrier is toxicity and harmful drug-drug interactions when multiple medicines are combined
- •Max frames the challenge as insufficiently “exquisite” engineering of molecules today
- •Better safety and reduced liver effects could allow more complex combinations
- •He describes an aspiration to modulate vectors of targets across pathways rather than single targets
Risk, regulation, and timelines: patient safety, differing risk appetites, and a five-year hope for better tools
The conversation turns to whether regulation overemphasizes safety and how risk-reward should be balanced. Max stresses patient safety as central, notes risk appetite varies by disease severity, and describes active engagement with regulators; he also expresses a hope that within five years, predictive biology and precise molecule design will “step change” significantly, even if development still takes years.
- •Patient safety is described as centrally important, with risk tolerance varying by indication
- •COVID and late-stage cancer are discussed as contexts with different risk appetites
- •Max says Isomorphic speaks with regulators globally and sees shared interest in better drugs for patients
- •He hopes AI-driven understanding will yield better success rates and safety profiles over time
- •He expects major improvements in predictive tools and precision design within about five years, while acknowledging timelines remain long
Longevity and healthspan: extending quality of life by tackling chronic afflictions and root causes
Elad asks how longevity-focused approaches relate to drug development, contrasting lifespan with healthspan and citing aging model observations. Max responds that impacting major diseases and chronic age-related declines could improve both longevity and quality of life, and he’s interested in efforts that address both symptoms and root causes of degradation.
- •Discussion distinguishes lifespan from healthspan and emphasizes late-life quality of life
- •Max points to chronic age-related afflictions that accumulate beyond major diseases
- •Examples include degradation in bones and muscles affecting mobility and mental health
- •He expresses interest in approaches targeting both afflictions and root causes of aging-related degradation
- •The episode closes with mutual appreciation and sign-off

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.