Lex Fridman PodcastDaphne Koller: Biomedicine and Machine Learning | Lex Fridman Podcast #93
At a glance
WHAT IT’S REALLY ABOUT
Daphne Koller on using machine learning to reinvent drug discovery
- Daphne Koller discusses how modern machine learning, coupled with new biological tools, can transform our understanding of disease and the drug discovery process. She explains the limits of current medical knowledge, especially for complex, heterogeneous diseases like Alzheimer’s and schizophrenia, and why traditional animal models often fail. Koller describes insitro’s strategy of building large, high-quality “disease-in-a-dish” datasets using stem cells, CRISPR, and advanced cellular measurements, then learning predictive models to find effective interventions. The conversation also touches on MOOCs and global education, AI safety and uncertainty, and her broader philosophy about impact, ethics, and the meaning of a good life.
IDEAS WORTH REMEMBERING
5 ideasWe understand only a small fraction of major diseases, and many are actually heterogeneous syndromes.
For conditions like Alzheimer’s and schizophrenia, Koller argues we are “closer to 0 than 80” in understanding mechanisms, and they likely comprise multiple distinct biological subtypes that only look similar clinically.
Animal models often fail because they mimic symptoms, not human mechanisms.
Mice rarely develop human diseases naturally; researchers engineer rough phenotypic copies that frequently don’t share causal pathways with human illness, so drugs that ‘work’ in mice fail in people.
Disease-in-a-dish models enable more human-relevant, data-rich experimentation.
By reprogramming human cells into induced pluripotent stem cells (iPSCs) and differentiating them into relevant cell types, then perturbing them (e.g., via CRISPR), insitro can observe disease-related cellular phenotypes directly tied to human genetics.
The real innovation is designing biology and experiments around machine learning needs.
Instead of treating ML as an afterthought on existing small datasets, insitro explicitly engineers large-scale, high-quality cellular and molecular data so powerful models can learn predictive representations useful for target discovery and intervention design.
New measurement technologies are turning ‘squishy’ biology into digital data.
Techniques like single-cell RNA sequencing and high-resolution microscopy generate quantitative, high-dimensional snapshots of cells and subcellular structure, providing rich inputs for ML to detect patterns, subtypes, and candidate interventions.
WORDS WORTH SAVING
5 quotes“We’ve been able to provide treatment for an increasingly large number [of diseases], but the number of things that you could actually define to be cures is actually not that large.”
— Daphne Koller
“Mice don’t get Alzheimer’s, they don’t get diabetes, they don’t get atherosclerosis, they don’t get autism or schizophrenia… and those cures don’t translate over to what happens in the human.”
— Daphne Koller
“What we are doing at insitro is actually flipping that around… putting [biological methods] together in brand new ways with the goal of creating datasets that machine learning can really be applied on productively.”
— Daphne Koller
“Machine learning algorithms today are really exquisitely good pattern recognizers in very specific problem domains… We’re nowhere close to the versatility and flexibility of even a human toddler.”
— Daphne Koller
“Our goal in life should be to make a dent in the universe… that I have left the world a better place than it was when I entered it.”
— Daphne Koller
High quality AI-generated summary created from speaker-labeled transcript.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome