Mice Don't Get Alzheimer's: Daphne Koller on Disease in a Dish

Lex Fridman opens with two far-out questions: will we one day cure all major diseases, and will we extend lifespan to the point of immortality? Koller refuses the "we will never" genre as hubris. "One day" is a very long time. Curing is still hard for a reason people skip: by the time you discover a disease, a lot of damage is already done. A cure at that stage assumes we can regenerate parts of the body back to their original state. That is a very challenging problem.

Humanity, she notes, has cured very few diseases. Treatments keep expanding; the set of things you can honestly call cures remains small. On a 0-to-100 scale of understanding fundamental mechanisms, the picture splits by disease. Biology is so complicated that a hundred is almost unavailable—new things keep appearing that nobody realized existed—but some diseases might sit in the seventies or eighties. The majority, she says, are really close to zero.

Alzheimer's is closer to zero than to 80. There are hypotheses, but they have not been sufficiently validated, and a growing number of people doubt that the traditional story explains what is going on. Schizophrenia is almost certainly near zero. Type 2 diabetes is a mix: insulin resistance is implicated and validated, yet many mechanisms remain opaque. The deeper claim is taxonomic. Alzheimer's, schizophrenia, and even type 2 diabetes are not really one disease. They are almost certainly heterogeneous collections of mechanisms that manifest in clinically similar ways, the same way breast cancer turned out to be many cellular routes to uncontrolled proliferation.

That understanding, Koller insists, has to precede any hunt for specific mechanisms. If a clinical label is actually ten diseases, a drug aimed at one mechanism will fail in the other nine. Trials recruit by phenotype, not by molecular subtype, then average the results and declare failure. The drug may have worked in the patients it was for.

Asked whether disease and longevity efforts overlap completely, Koller says the mechanisms certainly overlap without being identical. Other than childhood diseases, risk of contracting most diseases rises exponentially year by year from about age 40. That curve is the well-known link. Aging still happens without a named disease, and some disease mechanisms have little to do with aging.

At the cellular level she can already list wear and tear: DNA damage that accumulates as cells divide and repair fails to fully correct it; misfolded proteins that aggregate; inflammation those processes help drive. She is sure many mechanisms remain undiscovered. Fridman offers a philosophical tangent—that dying is a feature that lets new things grow, a kind of learning, both tragic and beautiful—and asks if she would choose immortality.

Immortal, she says, is a very long time, and not necessarily something she would want. What everyone aspires to is increased *healthspan*: more years of being healthy, active, and feeling as you did at 20. People deteriorate physically and mentally, and that is sad. A wonderful societal goal would be living to the biblical 120 in high quality of life. Whether the right number is 100, 120, or 150 is up for debate; on the timescale of the universe it is all pretty short.

Up to now, Koller does not think data and machine learning have played a significant role in understanding or eradicating disease. The datasets powerful methods need have not really existed—dribs and drabs, some interesting data science, not a revolution. The last few years are starting to change that: larger datasets, and, equally important, technologies that can produce data at scale.

Those tools have almost never been aimed, deliberately, at generating data for machine learning. Biologists produce data to drive scientific discovery; ML arrives as a by-product, a second stage of slightly fancier analysis. Insitro, the company she founded in 2018, flips the order. Take the repertoire of methods cell biologists and bioengineers have invented, put them together in new ways, and create datasets on which machine learning can actually build powerful predictive models for human health.

Fridman tries to summarize it as making data the primary goal. Koller corrects him: data is the means. The end goal is helping human health. The method they chose is predictive models, and those models, especially the more powerful ones, only work if the data has sufficient scale and quality. That is why they start from the data-generating end of the bench.

Her interest in health dates to the early 2000s, when machine-learning peers were still classifying the 20 Newsgroups dataset—bags of words from a concept that barely exists anymore. Biology datasets were more interesting technically and aspirationally, even when small. She wanted work that was societally useful, then grew more interested in biology itself, sometimes publishing papers with little ML in them.

Drug discovery became personal about twelve years before the interview, when her father died. He had an autoimmune disease that settled in his lungs. The doctors said there was only one thing they could do: give him prednisone. At one point a doctor proposed a lung biopsy to figure out which autoimmune disease he had. Koller asked whether that would change treatment. No. Prednisone was the only option. Friends who are rheumatologists later told her the FDA would never approve prednisone today; the ratio of side effects to benefit is probably not large enough.

Today, depending on the autoimmune disease, there are multiple drugs that can help, many of which did not exist twelve years ago. She thinks this is a golden time for making drugs that are much safer and more effective. What is lacking is enough understanding of biology and mechanism to know where to aim that engine. That, she says, is where machine learning can help. In August 2016 she had stepped away from Coursera with a blog post: it was time to turn to another critical challenge, machine learning applied to human health. This conversation was recorded before the COVID-19 outbreak.

Insitro's bet is what Koller calls a disease-in-a-dish model, for diseases where typical animal models have not worked. An animal model, often a mouse, is an external perturbation that creates a disease you then try to cure, hoping the similar-looking human disease will yield. The problem is that the way the animal is made sick often has nothing to do with how the human disease comes about. It is a copy of the phenotype, a copy of the clinical outcome, with quite different mechanisms.

Mice do not naturally get Alzheimer's, diabetes, atherosclerosis, autism, or schizophrenia. Cures in those animals usually fail to translate, and that is where most drugs fail. Disease-in-a-dish became possible only in the last five to ten years. You take a cell from a person—a skin cell, say—revert it to pluripotent stem-cell status, then differentiate it into the cell type you need: that person's neuron, cardiomyocyte, or hepatocyte, with that person's genetics.

If a genetic burden of disease would show up in that cell type, you can look at sick versus healthy cells and test interventions that might pull the unhealthy cell back. Curing cells is not curing people; a translatability gap remains. For diseases driven by human genetics that drive a cellular phenotype, there is reason to hope that reverting the cells where the disease begins will help more global clinical phenotypes. The backward step uses Yamanaka factors. Before that result, she says, very few people would have predicted it was possible. It was finicky at first and is now almost industrialized: contract research organizations will take a human sample and revert it, successfully a very good fraction of the time. The cells are not always perfect—they can retain bits of memory of being skin, or of environmental exposures—but by and large they are pretty good.

Reversal to stem cells is not yet a tens- or hundreds-of-thousands scale operation. The last time Koller looked, the world's stock of induced pluripotent stem cells sat somewhere between five and ten thousand, plus whatever sits in academic freezers. You cannot, at that point, generate iPS cells from a million people. Maybe you do not need to.

Once you have an iPS starting point, you can perturb it. One of the other miracles of the last decade is CRISPR gene editing. Take cells from a healthy person, introduce a known pathogenic mutation, and compare healthy and mutated cells one-to-one with everything else held constant. You can see specifically what that mutation does at the cellular level. More diversity is better—ethnic background matters—but you may not need a line from every patient with every disease.

People differ because their genotypes differ, and also because, for reasons not fully understood, some people's stem cells differentiate better than others. The difference that matters is disease burden: some of us carry more causative variants and fewer protective ones. Polygenic risk scores add those variants up. For diseases studied well enough, people in the highest decile of risk can be a factor of ten or twelve above the lowest decile. Genetics is not the full explanation, but there is real signal. There is even more signal, she believes, in the cells derived from those genomes. In principle all the signal is in the DNA; in practice our understanding of biology is so limited that watching the cell is a heck of a lot closer to the clinical outcome than staring at the sequence.

Fridman asks the outsider's question: cells are squishy; how do you plug them into a computer? Another revolution of the last ten years is the ability to measure cells quantitatively. When Koller started in biology in the late 1990s and early 2000s, microarrays could read the expression of every gene in a sample in one experiment. That is what revealed molecular subtypes of cancer. Until then it was "you have breast cancer"; the molecular data showed subtypes that look completely different at the level of gene activity.

Now single-cell RNA sequencing sequences the RNA of every gene in every cell. You literally count transcripts. That turns the squishy thing digital. Super-resolution microscopy, via digital reconstruction, can show subcellular structures, sometimes below the diffraction limit of light. Amazing scientists keep inventing new single-cell readouts. Those datasets can be used three ways. Backward: see a cellular phenotype and reason about which genes and pathways might cause it, using known biology. Forward: perturb a gene and ask whether the phenotype resembles patients', which would implicate that gene as causal. Insitro's path is the third: dump the large collection into machine learning, find subtypes that look similar in the clinic and distinct in the molecular data, then apply a drug or a CRISPR intervention and ask whether the disease state reverts toward healthy cells.

That third path is much less hypothesis-driven. It might surface interventions that are not the same targets everyone else is already staring at. Validation still follows. Koller will not promise a disease they will cure; she tries to deliver first, then promise. The approach is more likely to help when the disease has a strong genetic basis, when cellular models are reproducible enough to grow at low noise, and when the biology is contained in one or a few cell types you can put in a dish. Something broad, systemic, and distal is hard to cram in vitro. Bioengineers keep improving the systems: five years earlier, stem-cell-derived models barely existed and people worked in cancer cells, which are terrible models of most human biology. Now there are organoids—teeny multicellular organ systems, cerebral and liver and kidney. Brain organoids may be the coolest thing she has ever seen. Connecting them into multi-organ systems is still hard; in three or five years there will be disease models nobody can make today. Cancer-cell lines, passaged until their genomes are unstable, are not that future.

The conversation jumps to Coursera, which Koller co-founded with Andrew Ng. The origin sits in late-2000s Stanford experiments with online teaching for quality and scale. Ng led Stanford Engineering Everywhere, putting ten Stanford courses online as video lectures. Koller led an effort that broke courses into smaller units with embedded interactions, which university leaders liked as a path to what is now called the flipped classroom.

Those efforts fed one another and produced a burst of energy. In fall 2011 the first Stanford MOOCs launched. She did not invent the acronym—massive open online courses—and is not particularly fond of it, any more than "Big Bang" is a great name for the start of the universe, but it is what it is. There was no real publicity campaign, just a *New York Times* article that went viral. Within weeks, each course had about a hundred thousand students or more.

She remembers the conversation with Ng: there is a real need here. They were accomplished academics who could go back to writing papers. If they did that, this would not happen, and it seemed too important not to happen. They debated a Stanford project, a for-profit, a nonprofit, and chose the Coursera path, operating as a company from the beginning of 2012. The hunger, as she reads it, is not only for globalization. The world changed. Fifty years ago, college skills largely carried you through a career. Today many jobs need skills that did not exist when you went to college, and many jobs that existed then are dying. AI is part of that, not all of it. People need access to the skills they need *today*.

What they learned about teaching splits between ideas that translated back to campus and ideas specific to people learning inside a life. Short is better. People in the workforce cannot fit a 15-week semester. MIT's OpenCourseWare recorded classroom lectures of about an hour and a half; some people benefited, but it is not palatable if you have a job, three kids, and errands. They started with short modules and made them shorter. Fifteen minutes was still too long if you want to fit in while waiting in line for a child's doctor's appointment. Five to seven minutes works.

Courses should also be short, broken into units with a natural completion point so people feel close to finishing something meaningful and can come back for part two. Compressing content works because anyone who needs a slower pace can rewind. Engagement during the content matters, and faster feedback keeps people in. Micro-quizzes inside lectures and automatically graded assessments were an intuition she had going in, later validated by data. They even ran A/B tests on gender bias: a female instructor as role model can change the men-to-women balance in STEM courses, an experiment that would be hard on campus.

Crispness is not a free lunch in a lecture hall. People need time to absorb, pause, reflect, and practice; MOOCs give a chunk and then ask you to use it. Flipped classrooms use videos and exercises before class so face-to-face time can go to deeper, often group problem-solving. Any pedagogy beyond standing there and droning for an hour and fifteen minutes takes a heck of a lot more preparation. She does not think MOOCs will replace face-to-face teaching. Learning is often social; Coursera students formed study groups they did not have to form, and those learners did better. Recorded music did not replace live concerts. Continuing education, after high school and college are done, will move online because going back to school is not an option for most people.

For people who want to enter machine learning, her advice is to start, and not to skip the foundations of mathematics, statistics, and programming. She has watched people learn ML online or on campus without those foundations and then only turn the crank on existing models. They cannot innovate or adjust to the problem at hand; worse, they are sometimes just wrong and do not see the artifacts. Then find practical problems—at work, or Kaggle if you lack a workplace dataset—and find someone to solve them with, so you can bounce ideas and fix each other's mistakes. Practice.

Asked for the most beautiful idea in deep learning, Koller names two. The first is end-to-end training: from raw data all the way to the outcome you care about. It does not have to be a single monolithic blob; you can introduce building blocks trained on other tasks, and a blob in the middle is not ideal. What matters is that the whole chain can be trained. The second is learning a representation that, even if trained for another task, becomes a much faster starting point for a new one. That is reminiscent of what makes people successful learners. It is relatively new in ML and still underused relative to today's capabilities.

Was it surprising that neural nets work at all? She was surprised that, with enough data, you can find a meaningful representation in an exceedingly high-dimensional space. Early in her career, datasets were small, and she believed you needed a constrained, knowledge-rich search space to get a meaningful answer. That was true at the time. The open question is whether a completely knowledge-free approach will be the solution. It is not the solution today: a convnet for images is quite different from networks for language, speech, or biology. Some insight still goes into the structure. Will there be a universal learning machine? She does not know. She wonders whether insight always has to be injected somewhere, or whether that requirement can converge away.

A limitation of many models, including the ones she spent years on in probabilistic graphical models and Bayesian deep learning, is that they give an answer without a trustworthy sense of how much to believe it. Neural-net confidences are poorly calibrated: 0.8 versus 0.9 is not a real statement about reliability. Worse, the farther you move from the training data, the more wrong the network often is—and the more confident in that wrong answer. In medical diagnosis that can kill. Train on one patient population, put in an outlier the network has never seen, and it may deliver a completely incorrect diagnosis with supreme confidence. In driving you want the net to say it has no idea what the blob in the road is, and stop, rather than run over a pedestrian it does not recognize.

People are trying Bayesian deep learning, Gaussian processes, and ensembles trained on different data subsets or random starts, which are sometimes surprisingly good at expressing confidence. It is open research. Stuart Russell's point that more powerful systems should be full of self-doubt is, she says, true, especially when human life is at stake. The system needs to be able to say: I give up on this instance, because I have never seen something sufficiently like it.

On AGI, Koller splits feasibility, timelines, and protections. It is a long-standing dream, from the Turing test on. We are closer than 70 years ago and still very far. Today's algorithms are exquisitely good pattern recognizers in specific domains with enough training data. Move even a different version of the same problem, let alone a different problem, and they completely choke. "We're nowhere close to the versatility and flexibility of even a human toddler"—a child who can context-switch and solve different problems with a single brain.

She is not desperately worried about machines taking over the universe and killing people for power. Designing protections against human-level or superhuman intelligence, when we do not even see the skeleton of what that would look like, seems speculative. We can definitely, and already have, gotten into trouble with much dumber systems. The systems we build are increasingly complex and poorly understood; small changes have dramatic, unpredictable ripple effects. That is not unique to AI. The electric grid is really complicated. The software that runs financial markets is really complicated. Unanticipated feedback loops have already produced financial crashes. AI exacerbates the problem; the root is complexity and incomprehensibility.

Two engineering responses matter: better interpretability, and better tests of generalization. You will not test a car in every city, village, and weather. If you train on one set of conditions and it works on 50 or 100 quite different ones, you have some reason to believe the next 50 might work too. The other threat is abuse. Any powerful technology can be used badly; projectile missiles became guns, nuclear power became bombs. Gene editing and CRISPR are, to her, at least as dangerous as machine learning if used badly. You could create really nasty viruses. Fridman adds security attacks, surveillance, and killer drones. People who want to kill will find a way, she notes; the question is how we handle powerful tools.

On the long trend—fewer wars, less violence, more human rights—Koller is by and large optimistic. Most people mean well. That is not the same as being altruistic do-gooders, but most people do not set out to harm. Societies still have to build norms in which doing good and being well regarded by your peers are positively correlated. It is easy to create dysfunctional societies. Psychological experiments and, sadly, real-world events show people devolving into worlds where peer approval correlates with atrocious, often genocidal behavior.

She worries that some societies are not moving forward on that axis, that being a good person is not always what makes you look good to your peers. It is very easy to degenerate into a universe where it is okay to do really bad stuff and still have your peers think you are amazing. That, as much as any model, is a control problem for the technologies she spends her life on.

What gives her life meaning? When they were starting Coursera in fall 2011, Steve Jobs had just died, and the media was full of his quotes. One stuck, because it matched something she had already felt: our goal in life should be to make a dent in the universe. She hopes that on her deathbed she can point to ways she left the world better than she found it. She tells her kids that. The burden, she adds, is much greater for people born to privilege. She was not enormously wealthy, but she grew up in an educated family, with parents who loved her, a chance at a great education, and enough to eat—more privilege than the vast majority of humanity. Her kids have still more. For those who have that opportunity, using a life to make the world better is the point.