Lex Fridman PodcastDaphne Koller: Biomedicine and Machine Learning | Lex Fridman Podcast #93
CHAPTERS
- 0:00 – 2:30
Why Daphne Koller moved from Coursera to ML for human health
Lex introduces Daphne Koller’s background—Stanford professor, Coursera co-founder, and CEO of insitro—and frames the conversation as the early days of using machine learning to transform biomedicine. He also notes the recording predates COVID-19, underscoring the relevance of scalable drug discovery and health technologies.
- •Daphne’s roles: Stanford CS, Coursera co-founder, insitro founder/CEO
- •Machine learning as a data-driven lever for drug discovery and treatment development
- •Context note: recorded before COVID-19 but broadly relevant to pandemics
- •Lex sets a philosophical tone about disease, longevity, and impact
- 2:30 – 6:31
Can we cure all diseases? How little we understand many mechanisms
Daphne pushes back on sweeping predictions while explaining why “curing” disease is fundamentally hard—often damage is extensive by diagnosis and would require regeneration. She highlights that for many major diseases we may be close to zero in mechanistic understanding, and that several conditions are actually heterogeneous clusters rather than a single disease.
- •Avoiding hubris in long-term predictions; “one day” is a very long time
- •Cures are rare; treatments are more common than true cures
- •Many diseases have near-zero mechanistic understanding today
- •Alzheimer’s/schizophrenia likely heterogeneous rather than single mechanisms
- 6:31 – 10:10
Longevity as overlapping biology: aging, healthspan, and cellular wear-and-tear
The discussion shifts to longevity and how aging overlaps with disease risk, especially after ~40 as risks rise sharply for many conditions. Daphne emphasizes healthspan—staying healthy and functional longer—over immortality, and describes cellular processes that contribute to both aging and disease.
- •Disease risk often increases exponentially with age after ~40
- •Longevity and disease mechanisms partially overlap, but aren’t identical
- •Healthspan as a more meaningful target than immortality
- •Cellular contributors: DNA damage, misfolded proteins, inflammation
- 10:10 – 13:05
What ML has been missing in biomedicine: the right datasets (and insitro’s approach)
Daphne argues ML has had limited impact in biology largely because the necessary large, high-quality datasets didn’t exist. She describes insitro’s strategy: deliberately design and generate datasets using modern bioengineering tools so ML can build predictive models that meaningfully improve human health.
- •Historically: insufficient scale/quality of biological datasets for powerful ML
- •Recent changes: new technologies enabling data generation at scale
- •insitro flips the workflow: build datasets explicitly for predictive modeling
- •Data is a means; improving health is the end goal
- 13:05 – 16:25
Personal motivation: early bio interest and a family tragedy shaping drug discovery focus
Lex asks where Daphne’s health focus came from, and she recounts early dissatisfaction with uninspiring ML datasets and a pull toward biologically meaningful problems. She also shares how her father’s death from an autoimmune lung condition—where treatment options were limited—deepened her interest in better, safer, more targeted therapeutics.
- •Early 2000s: biology datasets felt more meaningful than classic ML benchmarks
- •Growing interest in biology beyond ML-centric work
- •Father’s autoimmune disease and limited treatment options (prednisone)
- •Belief we can now make safer/more effective drugs, but need better biology understanding
- 16:25 – 20:02
Disease-in-a-dish vs animal models: why translation fails and what iPS cells enable
Daphne explains traditional animal models, often mice, and why they frequently fail: the induced phenotype may not share the human disease mechanism. She contrasts this with disease-in-a-dish models enabled by induced pluripotent stem cells (iPSCs), where a person’s cells can be reprogrammed and differentiated to relevant human cell types to study genetics-driven disease phenotypes.
- •Animal models can mimic phenotype without matching human mechanism
- •Many mice don’t naturally get key human diseases (Alzheimer’s, schizophrenia, etc.)
- •iPSCs: reprogram adult cells to pluripotency, then differentiate into target cell types
- •Goal: observe disease phenotypes in human-genetics-derived cells and test interventions
- 20:02 – 23:50
Scaling and realism: Yamanaka reprogramming, CRISPR perturbations, and variability
The conversation dives into practical constraints: iPSC creation is increasingly industrialized but not yet at massive population scale. Daphne describes using CRISPR to introduce specific mutations for clean comparisons, and discusses sources of variability—genetic differences and differences in differentiation quality across lines.
- •Yamanaka-factor-style reprogramming is now more reliable but not perfect
- •Global iPSC line counts are still limited; scale is improving but constrained
- •CRISPR enables controlled “healthy vs mutated” comparisons in the same background
- •Variability: genetics plus differences in differentiation robustness across individuals
- 23:50 – 29:37
From genetics to cells to data: polygenic risk and measuring ‘squishy’ biology digitally
Daphne defines disease burden and polygenic risk scores as ways to quantify genetic predisposition across many small variants. She argues cell-derived phenotypes can be closer to clinical outcomes than genetics alone, especially given limited biological understanding. They then discuss measurement revolutions that turn cells into rich digital datasets.
- •Disease burden as aggregate genetic predisposition; polygenic risk scores
- •Genetics contains signal, but cellular phenotypes may be more predictive/meaningful
- •High-throughput measurement: microarrays → single-cell RNA-seq
- •Advanced microscopy (including super-resolution) yields quantitative subcellular structure data
- 29:37 – 33:24
How ML drives drug discovery in this paradigm: subtypes, perturbations, and prediction
Daphne outlines multiple ways to use cell data: hypothesis-driven backward reasoning, forward perturbation screens, and insitro’s pattern-finding ML approach. A central aim is discovering molecular subtypes and identifying interventions (drugs or gene edits) that shift diseased cells toward healthy states, then validating downstream.
- •Three approaches: backward biological inference, forward perturbation tests, ML pattern discovery
- •Identify molecular subtypes hidden behind similar clinical labels
- •Search for interventions that revert cellular disease signatures toward normal
- •Less hypothesis-bound discovery can uncover non-obvious therapeutic avenues
- 33:24 – 36:43
Which diseases fit disease-in-a-dish today: genetic basis, tractable cell types, organoids
Lex asks which diseases this can help; Daphne is cautious about promises and instead lists characteristics that make success more likely. She highlights genetic-driven diseases, robust/reproducible cellular models, and diseases localized to a small number of cell types, then points to organoids and multi-organ systems as the next frontier.
- •Better fit: strong genetic basis and clear cellular phenotype
- •Need reproducible, low-noise in vitro models at workable scale
- •Harder: systemic, multi-organ diseases that are difficult to recreate in a dish
- •Organoids (brain/liver/kidney/gut) and emerging multi-organoid connections expand tractability
- 36:43 – 41:51
The Coursera/MOOC origin story: Stanford experiments to global scale
The conversation pivots to education and the birth of MOOCs at Stanford, where Daphne and Andrew Ng explored online teaching for both quality and reach. The viral adoption of early Stanford MOOCs (100k+ learners) convinced them the demand was too important to ignore, leading to Coursera’s launch in 2012.
- •Late 2000s Stanford initiatives: online teaching quality + scale
- •Andrew Ng’s Stanford Engineering Everywhere and Daphne’s interactive modular approach
- •2011 Stanford MOOCs explode in popularity with minimal marketing
- •Decision process: Stanford effort vs nonprofit vs for-profit → Coursera
- 41:51 – 49:04
What makes online learning work: short modules, fast feedback, and flipped classrooms
Daphne distills lessons from Coursera’s experimentation: shorter is better at every level, and learners need flexibility and immediate feedback. She explains micro-quizzes, auto-graded assessments, and the flipped classroom model, noting these approaches demand far more instructor preparation than traditional lecturing.
- •Brevity matters: 5–7 minute videos often outperform longer lectures
- •Courses should be modular with natural completion points
- •Engagement via micro-quizzes and rapid feedback improves persistence
- •Flipped classroom: content before class, deeper problem-solving in-person; higher prep burden
- 49:04 – 55:08
Advice for learning AI and the most beautiful deep learning ideas
Daphne advises aspiring ML practitioners to build strong foundations (math, stats, programming), then learn ML and practice on real problems, ideally with collaborators. She names end-to-end learning and representation/transfer learning as especially powerful ideas, while reflecting on how data scale changed what’s possible in high-dimensional learning.
- •Start learning, but don’t skip fundamentals (math/stats/programming)
- •Move from theory to practice via real workplace problems or Kaggle
- •End-to-end training as a foundational deep learning paradigm
- •Transfer learning/representation learning as underutilized and human-like in spirit
- 55:08 – 1:06:35
Uncertainty, robustness, and AI safety: calibrated confidence and risks from ‘dumb’ systems
Daphne discusses the dangers of poorly calibrated ML confidence, especially in medicine and autonomous driving, and surveys techniques like Bayesian approaches and ensembles. She then tackles AGI: she views it as far away, but emphasizes real near-term risks from complex, poorly understood systems, misuse, and societal deployment choices.
- •Neural networks can be confidently wrong; calibration is critical
- •Medical diagnosis and self-driving illustrate high-stakes failure modes
- •Methods: Bayesian deep learning, Gaussian processes, ensembles; still open research
- •AGI seen as distant; nearer risks: system complexity, robustness testing, misuse (surveillance, weapons), and parallels to CRISPR dangers
- 1:06:35 – 1:12:03
Human nature, social norms, and a personal meaning-of-life framework
Lex asks whether people are fundamentally good; Daphne is broadly optimistic but warns that social norms can drift toward rewarding harmful behavior. She closes on meaning and purpose: “making a dent in the universe,” improving the world—especially as a responsibility of privilege—and teaching that ethos to her children.
- •Most people mean well, but societies can incentivize harmful behavior
- •Importance of social norms aligning status with doing good
- •Meaning of life as leaving the world better than you found it
- •Privilege increases responsibility to contribute positively
