Co-Founder, Coursera
insitro Founder and CEO Daphne Koller at Semafor Tech: First Principles
insitro Founder and CEO Daphne Koller joins Semafor tech editor Reed Albergotti at Semafor Tech: First Principles, hosted June ...
About Daphne Koller
Daphne Koller, CEO of insitro and co-founder of Coursera, has been speaking about the application of artificial intelligence to biology and drug discovery. At a Semafor Tech event in June 2026, she argued that the field is at an inflection point, citing stronger computational and experimental tools than when she began her career. She discussed insitro's work on an "ALS disease axis" identified by printing and phenotyping over 12 billion motor neurons. Koller also advocated for a "data commons" for healthcare data, comparing it to organ donation, and expressed skepticism about extending maximum human lifespan past 130, though she said extending health span is a realistic goal. In other appearances, Koller reflected on her career trajectory from academia to industry, stating that building a product people will use is necessary for real-world impact. She described the current era as one where "the currency of success is imagination, perseverance, and judgment" in deciding what to build. Koller also noted that while many scientific discoveries originate in the U.S., China executes on them faster due to fewer bureaucratic hurdles. She described probabilistic graphical models as a framework for modeling causal relationships in complex domains, and said that as AI systems become more agentic, understanding causality will become increasingly important.
Transcript (40 segments)
✨ AI-enhanced transcript with speaker attribution
H
Host0:00
We welcome Daphne Koller of Insitro and Semafore's Reid Al Bagoati to the stage.
Daphne.
D
Daphne Koller0:08
Thank you.
H
Host0:11
Daphne, it's good to see you. Thanks for being here.
D
Daphne Koller0:14
Thank you for inviting me.
H
Host0:15
Yeah. So, you're one of the early pioneers in the use of AI in therapeutics and science. You now have a company. This is a hot area now. So, my question is: are we at some sort of inflection point right now, or are people just sort of paying attention?
D
Daphne Koller0:34
I think it's both. People are absolutely paying attention, and even if you look at the big foundation model companies out there, everyone's like the killer app for artificial general intelligence or artificial super intelligence is going to be science. So people are paying attention, but I think there is a sense in which we are really at the cusp of a transformation because the tools that we now have available to us on both the computational side and the experimental side are so much stronger than what we had when I started my career in this space.
H
Host1:10
Yeah. I want to outline the problem here, because I could get a laptop and use AlphaFold to design a drug today. That doesn't mean it's going to become a drug, right? It doesn't mean people's lives are going to be saved.
D
Daphne Koller1:25
I'm going to modify what you said and say that you can use AlphaFold or one of its children to design a protein. Whether that protein is a drug is a question that has multiple ramifications. Can you deliver it safely and effectively to a human is maybe table stakes, but the bigger question, the one we've set out to solve, is: even if you can take that protein, put it in a person, and it performs a certain biological function, does it do anything for the disease? For the vast majority of diseases, we have no idea where to intervene. So we can't even tell you what type of protein you need to design.
H
Host2:06
It's a very daunting space. So maybe you could help draw a line: what is AI useful for and what is it not useful for?
D
Daphne Koller2:16
In this space.
H
Host2:17
Yes.
D
Daphne Koller2:18
Let me change the question slightly. I can tell you what people are using it for and what I think they should be using it for and what we're using it for. There's a lot of work on exactly what you said: designing a molecule that performs a particular function, binds in this way to a target. That can be a small molecule or a biologic. We've gotten far better at designing those molecules and cutting the time from years to months to weeks, and arguably some might say hours. That's awesome. It shortens timelines, which is nice, but it doesn't answer the fundamental question of what is the right intervention. It's like human judgment. We have AI agents that can do exactly what you tell them, but knowing what to tell them is the fundamental problem that's currently rate limiting on humans and, in our case, rate limiting on collecting the kind of data that really disentangles the underlying biology of disease.
H
Host3:18
Yeah. One of the things you're doing at Insitro is focusing on ALS, among other things.
D
Daphne Koller3:25
Among other things.
H
Host3:25
We could talk about brown fat, but ALS is a big one. I'm curious how you chose that, because choosing the right target is the big battle.
D
Daphne Koller3:38
Absolutely. We picked ALS due to a set of considerations. The unmet need is enormous: people die within 3 to 5 years in a horrible death, and there's no therapeutic that meaningfully moves the needle. That's an obvious reason. But we also felt we had a discovery approach that could unlock differentiated insights for this disease. We were able to interrogate the biology of healthy and diseased motor neurons and revert a disease motor neuron back to a healthy state, where the definitions of healthy and diseased and the trajectory needed were revealed to us by AI methods looking at billions of these motor neurons. The opportunity was where we could bring disproportionate value using our discovery approach.
H
Host4:45
Right. One of the things is there was a data set out there, right?
D
Daphne Koller4:49
No, there wasn't. We made it.
H
Host4:50
You made it. How did you make the data set?
D
Daphne Koller4:53
We spent much of our journey at Insitro building the capability to print biology at scale. We printed 22 billion cells and counting, measuring them in different ways, perturbing them in different ways, so that we can build what hasn't been built before: an AI model for human causal biology. Causality is really important because you can collect observational data sets all day long and get the passengers that come along for the ride in the disease. But causality is critical when you're intervening in a system. So we are building the data set that will allow us to create an emergent model of human causal biology and tell us the intervention points that will be impactful.
H
Host5:40
It's now your proprietary data. So I'd love to know: we have this new power to analyze this data, yet other than a few countries like the UK Biobank, maybe the UAE, maybe China, they're not doing much that's ambitious around data.
D
Daphne Koller6:04
We are not, and honestly, it's very sad for me. We are able at this point to print large amounts of cellular data, and that can absolutely reveal biological mechanism. But at the end of the day, you're not curing cells; you need to cure people. The bridge to translatability is what it takes to move the needle in a human being. For that, you need human data. You correctly said UK Biobank is a landmark data set. There are a few others. In the United States, we have done precious little to create a large national resource that will empower that type of scientific discovery. I find it shocking that I can go to the DMV today and check a box to have my organs donated to science, but I can't check a box to have my healthcare data donated to science. Why is it that patients and their families don't have the agency to say there is a data commons, appropriately safeguarded with anonymity, so I can donate my data to empower discovery?
H
Host7:11
On that privacy question, there are people who say it doesn't matter if it's anonymous; someone will find a way to reverse engineer the data and de-anonymize it. Do you think those are legitimate concerns?
D
Daphne Koller7:24
It depends on exactly what you're releasing. There are parts of your data that I would challenge anybody to try and understand who that person is. If you just release labs, what are you going to do with that? But it's still very valuable. Other things I agree have more privacy considerations. But no one kicks up a fuss when credit card companies give out our entire record of every purchase, which is a lot more revealing for many people than the healthcare issues they may have been born with, which is not on them.
H
Host8:04
What would be, if you could snap your fingers and be emperor of the world, what would you actually collect from people? DNA, epigenetic data?
D
Daphne Koller8:16
I would start with what we already have. There are millions and hundreds of millions of people from whom we've collected tons of healthcare data throughout their life journey. I would start with that, appropriately consented, and use AI to make sense of longitudinal patient journeys and how that has impacted downstream consequences. Then, staying rational because I'm not the emperor, I would identify some number of individuals with particularly interesting healthcare journeys, deeply phenotype them, measure all sorts of aspects like epigenetics, imaging, to relate the clinical consequences from the large population to the biology we measure in that smaller number. That's what I would do.
H
Host9:18
Yeah. We mentioned some of those countries. I brought up China. They've been racing ahead in biotech and made news today on brain computer interfaces. Are we still ahead, or is China catching up?
D
Daphne Koller9:36
I think in some ways we are in the most cutting-edge of biology discovery. Most of those discoveries are still initiated in the US and then very rapidly taken up and capitalized on by the Chinese. They can execute way faster because they have developed a machine that allows them to experiment at scale in the lab efficiently, and more recently in humans in clinical trials. People assume they're doing things unconsented and unethical, but that's not true. They've put in place reasonable ethical considerations. They just removed all the bureaucratic crap that makes this slow in the US. They can rapidly take a program into the clinic, collect data from humans, which is the only model system that translates to humans, and use that to improve their next program. For us, that journey takes 6 years.
H
Host10:45
Could we do that in the US?
D
Daphne Koller10:46
With the right willpower and capabilities, we absolutely could. I would argue that AI also creates opportunities to accelerate some of that clinical development in ways the Chinese haven't adopted yet. For example, if we had my magic emperor database of anonymized US healthcare records, I could identify individuals most likely to benefit from a particular clinical trial. That would not only allow fast recruitment but also democratize clinical trial access to populations that currently don't benefit because they don't live near an academic medical center. So participation in clinical trials would become available to rural America and disenfranchised populations.
H
Host11:46
What's stopping us from doing that now?
D
Daphne Koller11:49
It's a government-level decision.
H
Host11:52
You've talked to people in government?
D
Daphne Koller11:54
I have tried to make that case, and hopefully we're starting to get some people to listen. I would also point out that the US government pays for healthcare of a large fraction of the population via Medicare and Medicaid. So does that give us the right to have an expectation that that data could be used, with appropriate consents and opt-out options? Do we not want to make the discovery of novel medicines faster and more effective? Let me share some statistics: of all named diseases, only 22% have any approved drug. Half of those are common diseases. So there is enormous unmet need. Moreover, the number of novel mechanisms pushed into the clinic annually has gone from 100 a decade ago to 30 now. The innovation to address unmet need is decreasing precipitously. How is that going to help all those sick people?
H
Host13:22
One interesting thing happening with GLP-1s is that they exist outside the healthcare system; people pay out of pocket. Do you think the economics are changing? Does that increase entrepreneurship?
D
Daphne Koller13:47
I don't think the GLP-1 model can be replicated for diseases that aren't mass market where people are incentivized to pay out of pocket to look great at the beach. For many diseases, most people cannot afford to pay out of pocket, nor should they. But the way we've structured healthcare economics, you pay for drugs out of pocket, while things that bring less benefit, like hospital visits, are covered. So some of the most effective interventions, drugs, are viewed as the culprit because of the incentive system. There's absolutely better systems we could design for equitable access to life-transforming interventions.
H
Host15:00
As we run out of time, one question in the longevity space: people talk about peptides and all that. Are we on the verge of a change in that space, or is it all BS?
D
Daphne Koller15:23
I don't think it's all BS. There are some very interesting interventions that could extend health span. What I'm not convinced of yet, because no one has shown it, is pushing past the natural boundaries of maximum lifespan. No one has created anything suggesting you can live past 130. But extending health span is really cool. We didn't talk about our brown adipose tissue target, which appears to activate white fat and turn it into metabolically active, healthier brown fat. That sits upstream of many diseases of aging like cardiovascular disease, kidney disease, liver disease, obesity, diabetes. More and more of those targets will emerge, and they'll be better than some weird Chinese peptide from Etsy.
H
Host16:26
That's all the time we have. Maximum lifespan will be relevant in later conversations. Daphne, thank you so much.