Koller, Daphne

professor of computer science at Stanford; co-founder of Coursera; founder and CEO of Insitro, a machine learning and biomedicine company

Daphne Koller: Biomedicine and Machine Learning

Lex Fridman Podcast · May 5, 2020

• Daphne Koller
• Insitro
• machine learning and medicine
• disease-in-a-dish models
• induced pluripotent stem cells
• CRISPR
• Coursera
• MOOCs
• drug discovery
• AI uncertainty

Core ideas

Koller resists making grand predictions about curing "all disease" or achieving immortality, noting that by the time most diseases are diagnosed, significant damage has already occurred, and that true cures (versus treatments) remain rare. She estimates progress unevenly across diseases: some conditions like certain cancers are reasonably well understood (she'd put understanding in the 70s–80s on a 0–100 scale), while Alzheimer's, schizophrenia, and to a lesser extent type 2 diabetes are "closer to zero" — partly because, like breast cancer (now understood as a heterogeneous collection of distinct cellular mechanisms producing similar clinical symptoms), these are almost certainly not single diseases but bundles of distinct underlying mechanisms that haven't yet been disentangled. She connects aging and disease as overlapping but not identical processes (disease risk rises roughly exponentially after age 40, driven by mechanisms like DNA damage accumulation and misfolded protein aggregation), and expresses a personal aspiration for an extended "health span" — living to something like a healthy 120 — rather than literal immortality, while resisting a firm view on the ideal target age.

Koller traces her focus on drug discovery partly to watching her father die of an autoimmune lung disease 12 years ago, when doctors had only prednisone to offer (a drug she says wouldn't meet today's FDA safety/efficacy bar) — she frames the current era as a "golden time" in drug discovery where safer, more effective drugs are technically achievable, but where the bottleneck is insufficient mechanistic understanding of biology to know where to aim them, which is where she sees machine learning's opportunity. Insitro's core strategy, as she describes it, inverts the usual order: rather than applying ML to whatever biological data happens to already exist, the company uses bioengineering techniques specifically to generate large, high-quality datasets that machine learning can act on productively — treating data generation itself as an engineering problem in service of building predictive models for human health.

The technical core of the conversation covers "disease-in-a-dish" models as a response to animal models' key weakness: mice don't naturally get Alzheimer's, diabetes, atherosclerosis, autism, or schizophrenia, so artificially inducing disease-like phenotypes in them often fails to capture the actual human disease mechanism, which is a major reason drugs that work in mice fail in humans. Enabled by Yamanaka-factor reprogramming (converting an ordinary cell like a skin cell into an induced pluripotent stem cell, which Koller still finds astonishing — "before that happened, I think very few people would have predicted that to be possible"), researchers can now differentiate a patient's own cells into neurons, cardiomyocytes, or other cell types carrying that patient's genetics, allowing side-by-side comparison of "sick-looking" versus healthy cells — and, using CRISPR gene editing, introduce a single known pathogenic mutation into an otherwise identical cell line to isolate its specific cellular effects. She notes the global supply of induced pluripotent stem cell lines is still small (roughly 5,000–10,000), but argues this may be less limiting than it seems since CRISPR-based mutation studies can substitute for needing IPS cells from every possible patient, and that polygenic risk scores already show substantial (often 10x+) differences in disease risk between people at the extremes of genetic risk distributions, providing real machine-learning signal even before cellular data is added.

On measurement technology, Koller highlights single-cell RNA sequencing (quantifying gene activity/expression at individual-cell resolution — "turning that squishy thing into something that's digital") and super-resolution microscopy (digitally reconstructing sub-cellular structures below the diffraction limit of light) as the recent enabling breakthroughs that make disease-in-a-dish models tractable for machine learning, following in a lineage that goes back to microarray gene-expression profiling revealing molecular subtypes of diseases like breast cancer. She describes Insitro's actual machine learning approach as neither purely backward (working from an observed cellular phenotype to infer causal genes) nor purely forward (perturbing a candidate gene and checking for a disease-like phenotype), but a third path: using ML to discover patterns and subtypes across large cellular datasets that share similar clinical outcomes but distinct molecular signatures, then testing whether specific interventions (drugs or CRISPR edits) can revert a disease subtype's cells toward a healthy-looking state — a much less hypothesis-driven discovery method than traditional approaches. She's deliberately cautious about promising specific disease cures, but outlines the characteristics that make a disease more tractable for this approach: strong genetic basis, reproducible and robust cellular models, and disease mechanisms contained within one or a small number of cell types rather than distributed systemically — while noting that emerging technologies like multicellular organoids (cerebral, liver, kidney) and even connected multi-organ systems are rapidly expanding what's tractable.

Turning to Coursera and online education, Koller recounts co-founding it with Andrew Ng after Stanford's early free online course experiments (including Ng's "engineering everywhere" video lecture project and her own more interactive course redesign) attracted roughly 100,000 students within weeks with no real marketing push, convincing them the effort was too important to abandon. Key lessons she cites: shorter is dramatically better (courses moved from MIT OpenCourseWare-style 90-minute recorded lectures to 15-minute, then 5–7-minute video segments to fit around real adult schedules), breaking 15-week courses into shorter modules with natural completion points sustains motivation, embedded low-stakes quizzes with immediate feedback improve engagement, and — validated via online A/B testing infeasible on a physical campus — a female instructor role model measurably shifted gender balance in STEM course enrollment. She argues effective non-traditional teaching (both MOOCs and face-to-face "flipped classroom" models) actually requires substantially more preparation than traditional lecturing, and predicts MOOCs will supplement rather than replace in-person education (citing organically-formed learner study groups correlating with higher success rates), becoming most valuable for continuing/lifelong education given how fast job-relevant skills now change.

On AI more broadly, Koller identifies end-to-end training and transfer/representation learning as deep learning's most beautiful and surprising ideas, expressing genuine surprise that meaningful representations can be found in extremely high-dimensional spaces given large enough data, while noting current architectures still require domain-specific inductive biases (vision vs. language vs. speech networks differ substantially), leaving open whether a truly universal, knowledge-free learning architecture is achievable. She considers poorly calibrated uncertainty a serious limitation of current ML — models can be simultaneously wrong and highly confident, especially on out-of-distribution inputs, which she calls a serious risk in high-stakes domains like medical diagnosis or self-driving cars, and points to Bayesian deep learning, Gaussian processes, and model ensembles as active but unresolved research directions. On AGI, she considers today's systems still far short of even a human toddler's contextual flexibility, and is not worried about near-term superintelligence risk — but argues real, present risks come from the growing complexity and poor interpretability of powerful systems generally (comparing this to unpredictable ripple effects in the electrical grid or financial markets), from insufficient testing of generalizability across untested conditions, and from potential misuse of powerful technology (drawing an explicit parallel to gene editing/CRISPR as comparably dangerous if misused as machine learning). She closes expressing measured optimism that most people "mean well," while emphasizing the importance of maintaining social norms where being perceived as good is actually correlated with being good, warning this correlation can break down catastrophically (citing historical examples of societies where atrocious behavior was socially rewarded).

Excerpts