# Ep213: Daphne Koller on The Moment in AI Drug Discovery

- **Show:** The Long Run with Luke Timmerman
- **Published:** 3 Nov 2026
- **Source:** https://soundcloud.com/longrunpodcast/ep213-daphne-koller-on-the
- **Transcript:** https://podcasts.opennash.com/shows/the-long-run/ep213-daphne-koller-on-the-moment-in-ai-drug-discovery-1d17c7/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Daphne Koller, CEO of Insitro, argues that AI will transform drug discovery but not as fast as the hype claims. 'Curing all diseases in five years' is physically impossible. Biology data is scarce and messy, so useful AI requires collecting purpose-built data in labs and human cohorts, then pairing specialized models with general-purpose language models.

The central tension is between the huge long-term opportunity and the slow, hard reality of proof. Insitro has nine programs, with fatty liver, ALS and eye disease drugs heading to the clinic in 2027 and 2028. Koller says AI speeds the middle step of making drugs, but choosing the right target and running trials still depend on deep human biology, and human data remains the biggest gap.

## Key insights

1. **Biology lacks the data that powers AI elsewhere** Koller says public data is a tiny, often non-reproducible fraction of what is needed, with 50 to 80 percent of literature not reproducible. So Insitro built a wet lab 'data factory' because there is no internet-scale biology dataset to scrape.
2. **Two-sided platform links cells to human genetics** Insitro runs perturbation experiments in human cells and pairs them with large human cohort data. Natural genetic variation supplies cause-and-effect evidence that bridges the two, helping show what may translate to patients.
3. **Specialized models plus LLM agents work best** Insitro adapts purpose-built models for microscopy and MRI rather than feeding raw data to chatbots. LLM agents handle quality control, experimental design and literature integration, cutting design work from six to nine months to eight weeks.
4. **Insitro's pipeline reaches clinic in 2027–2028** The MASH siRNA drug targeting IRS-1 is expected to start patients in early 2027. ALS and dry AMD programs are aimed at early 2028, and the ALS antisense drug targets TDP-43 pathology seen in 97 percent of patients.
5. **AI's biggest near-term win is making drugs** Koller sees the most legitimate progress in turning a chosen target into a drug quickly. She says target selection and clinical trial design still need deeper disease understanding, where she sees less progress, and human data is the scarcest resource.

## Questions this episode answers

### Can AI cure all diseases in the next five years?

No. Koller calls that physically impossible because the data must be collected, drugs made and clinical trials run, and there aren't enough trial sites. She does think that in five to ten years there could be a path forward for a good fraction of diseases now seen as impossible.

### Why can't AI companies just train on existing biology data like chatbots do on the internet?

Koller says the existing data is a tiny fraction of what is needed, was not collected for AI, and is full of artifacts and irreproducible results. Human diseases like neurodegeneration often can't be studied in other species, so the data was never collected.

### What is Insitro's lead drug and how does it work?

The lead program treats MASH, a fatty liver disease, with a liver-targeted siRNA (a drug that silences a gene) aimed at IRS-1. Koller says IRS-1 regulates both liver fat and scarring, and first patients are expected in early 2027.

### Why does Koller think Insitro's ALS drug could be disease-modifying?

In lab-grown motor neurons with TDP-43 pathology, the drug reduced toxic RNA errors and restored healthy proteins. Cut-off nerve fibers also regrew, which Koller called the key functional recovery, though human trials are still ahead.

### Do GLP-1 drugs make new fatty liver treatments unnecessary?

Koller says no. GLP-1 effects on liver fat wane quickly and are modest, and they don't help people with lean MASH. She says an siRNA lasting three to six months could also be friendlier than weekly injections.

## Facts

- **Luke Timmerman** [0:00:01]: "Insitro was founded in 2018 and has secured $750 million in funding from venture capital firms and through partnerships with big drug makers"
- **Daphne Koller** [0:05:36]: "the fraction of diseases that have any FDA-approved therapeutic, even if it's addressing purely symptoms, depending on whose estimates you believe, it's somewhere between 9 % to 22%"
- **Daphne Koller** [0:09:45]: "our success rate of a treatment that goes into the clinic is stubbornly shy of 10%."
- **Daphne Koller** [0:14:11]: "Somewhere between 50 to 80% of scientific literature is not reproducible."
- **Daphne Koller** [0:22:29]: "we have at this point onboarded, I think, 28 petabytes of data."
- **Daphne Koller** [0:22:29]: "we also have somewhere between three and four of those petabytes are human clinical data."
- **Daphne Koller** [0:22:58]: "our partnership with the Moorfields Hospital in the UK that has one of the world's largest collections of OCT data, which is a form of eye scans."
- **Daphne Koller** [0:23:55]: "our really exciting partnership with Lilly that allowed us to predict in vivo pharmacokinetic properties for small molecules."
- **Daphne Koller** [0:26:46]: "took a process that used to take our scientists like six to nine months and shrunk it down to, you know, eight weeks."
- **Daphne Koller** [0:29:01]: "So we have three programs that are heading into the clinic in the next 18 months."
- **Daphne Koller** [0:29:01]: "they are exiting the liver space and focusing on virology and oncology"
- **Daphne Koller** [0:29:47]: "we're going to file this year and first patient in will be early in 27."
- **Daphne Koller** [0:31:50]: "This one is IRS-1, and the genetics are incredibly strong for this."
- **Daphne Koller** [0:32:09]: "A liver-targeted siRNA, which is actually really critical because you would not want to inhibit IRS-1 systemically."
- **Daphne Koller** [0:33:04]: "We filed a patent in order to protect the composition of matter, And for an oligo therapeutic, the sequence that you are disclosing in the patent tells you the target."
- **Daphne Koller** [0:38:13]: "SIRNAs have very long durability. This one is easily three to six months injection"
- **Daphne Koller** [0:21:20]: "we have been, we believe, able to find not one but three programs that truly are disease-modifying in the 97 % of ALS patients that have what is called TDP-43 pathology"
- **Daphne Koller** [0:39:52]: "TDP-43 that's supposed to be doing a really critical function is not doing that function anymore."
- **Daphne Koller** [0:47:07]: "BMS elected to pursue the first of those targets, ALS-1, as a small molecule."
- **Daphne Koller** [0:48:17]: "then we hope to get that into the clinic in early 2028."
- **Daphne Koller** [0:44:14]: "This is a biogen drug specifically against SOD1 patients"
- **Daphne Koller** [0:49:38]: "he worked when he was at Amgen on Meritide, which is their incretin"
- **Daphne Koller** [0:52:04]: "if the MRI was collected in winter, you have more brown fat."
- **Daphne Koller** [0:53:01]: "that mouse lost 25 % of its fat mass. It didn't lose any lean mass."
- **Daphne Koller** [0:54:04]: "It's the largest cause of blindness in adults over 50."
- **Daphne Koller** [0:54:57]: "we were able to train the world's largest foundation model for oct data"
- **Daphne Koller** [1:01:56]: "Protein folding is actually closer to physics than it is to biology."
- **Daphne Koller** [1:03:08]: "we've never measured any of those things before in 500 ,000 individuals."

## Transcript

**[0:00:01] Luke Timmerman:** Welcome to The Long Run. This is a podcast for biotech adventurers. I'm your host, Luke Timmerman. Today's guest is Daphne Koller. Daphne is the founder and CEO of South San Francisco-based Insitro. It's one of what you could call the first-generation companies that uses AI to speed up and improve drug discovery. Insitro was founded in 2018 and has secured $750 million in funding from venture capital firms and through partnerships with big drug makers like Bristamire Squibb, AstraZeneca, and Gilead Sciences.

**[0:00:39] Luke Timmerman:** Insitro has been collecting and curating biological data to get to the point where AI can be useful enough to even approach the most important hurdle in R &D, that being human clinical trials. In this conversation, Daphne offers a reality check on what AI can and can't do at this moment for the discovery of new medicines. I think this conversation is worth listening to for a balanced view from someone in the trenches who is both optimistic and realistic.

**[0:01:07] Luke Timmerman:** She goes into specific detail about programs that her team is developing for fatty liver disease, ALS, and brown fat-driven markers with implications for cardiometabolic disease. Now before we get started, a word from the sponsor of the show, AlphaSense. Anyone who's tried to stay current on a fast-moving therapeutic area knows the real problem isn't finding information. It's that there's too much of it and it changes every week. AlphaSense just launched Super Analyst, an always-on agent that works like a power user inside the platform on your behalf.

**[0:01:41] Luke Timmerman:** It draws on AlphaSense's full life sciences content universe. Clinical trial data, peer-reviewed literature, regulatory filings, conference presentations, and expert transcripts with researchers, clinicians, and executives, and it keeps monitoring your space even after you've closed your laptop. It's designed to act like an extension of your research team. It remembers the programs and competitors you're tracking, follows your process, and applies it automatically to new data as it comes in. And every output links back to its original source, so when you're citing a trial result, you can see exactly where the number came from.

**[0:02:14] Luke Timmerman:** Take the task you've been meaning to get to, tracking a competitor's trial program or flagging the next regulatory decision before it hits, and hand it to superanalyst. Start your free trial at alpha-sense.com slash the long run. That's alpha-sense.com slash the long run. And quick question. When was the last time a CRO told you what something costs before you sign an NDA, sat through two scoping calls, and waited a week for a scope of work?

**[0:02:46] Luke Timmerman:** In bioanalysis, pricing gets treated like a closely guarded secret. Dash Bio thinks that's backward. It publishes pricing publicly on its website and even has a calculator that shows how the price changes if you redesign your study. You can see what your study costs before you talk to a human being. No surprise, line items, no it depends, no games. Dash was built to be the bioanalysis company that's actually straightforward to work with. Transparent pricing, guaranteed outcomes, and data in days, not months.

**[0:03:17] Luke Timmerman:** If you're tired of pricing that feels like a negotiation every single time, run your bioanalysis with Dash. Visit dash.bio slash pricing and see exactly how much your next bioanalysis study will cost. Now please enjoy this conversation with Daphne Koller on the long run. Daphne Koller, welcome to The Long Run.

**[0:03:40] Daphne Koller:** Luke, very nice to see you again.

**[0:03:43] Luke Timmerman:** So, Daphne, it's hard to believe that it was six years ago that you were a guest on this podcast in the early days of Insitro. A lot has happened in artificial intelligence broadly and with artificial intelligence for drug discovery in the years since.

**[0:04:00] Daphne Koller:** Indeed, it's hard to believe we're living in the same world.

**[0:04:05] Luke Timmerman:** And, you know, one thing that I've noticed recently that I wanted to follow up and ask you about and have you on the show is about there's so much discussion in the public about AI and its potential to cure cancer, cure diseases, cure all diseases, based on what some at the leading foundation model companies are saying. You have been working in this area for a long time and have a deep understanding of what the AI can and can't do.

**[0:04:37] Luke Timmerman:** Can you provide me and the listeners of this show with a little bit of a reality check, a balanced view of where things are and where you think they could go?

**[0:04:47] Daphne Koller:** So, first of all, I think it's important to be very thoughtful in how we use the word cure. Cure in the sense of you have an intervention and now you are not sick anymore. We have cured a precious few diseases of any type as a medical industry ever. Some childhood cancers, infectious diseases, and pretty much very little else. And so I think we should be just thoughtful in how we use that word. But even if we put that aside and we set the bar as, you know, we took something that used to be a dire condition and turned it into a chronic manageable condition, which is a lower and more achievable bar, there's still, I think, a tremendous amount of

**[0:05:36] Daphne Koller:** hubris in that statement. because for the number of diseases that have any FDA-approved therapeutic, sorry, the fraction of diseases that have any FDA-approved therapeutic, even if it's addressing purely symptoms, depending on whose estimates you believe, it's somewhere between 9 % to 22%, which means that the vast majority of diseases we're currently offering patients absolutely nothing. I would then go and say that when you take a step beyond that and ask, well, for what fraction of those that are untreated, We know what we need to do and we just haven't been able to make the molecule.

**[0:06:13] Daphne Koller:** The answer is also a diminishingly small fraction. And even for those where we think we know the answer, often, unfortunately, we turn out to be wrong. So we need to stand in front of the complexity of human biology with the appropriate level of humility and recognition that we only understand a very small amount of the beauty and complexity of human physiology. And there's a lot that we still need to figure out. And I think that's a really exciting journey and one where AI will be incredibly important and frankly, the only technology that can get us from where we are today to a sufficiently actionable understanding that we can actually start to treat some of these diseases.

**[0:07:00] Daphne Koller:** That's going to take time. It's going to take time to collect the data, most of which we do not, haven't even begun to collect. It's going to take time to interpret and figure out the actionable interventions. It's going to take time to then make the drugs. And it's going to take time to actually prove them out in the clinic because we are nowhere close to being able to do an in silico clinical trial that is in any way definitive about what things will do in a human.

**[0:07:26] Daphne Koller:** So AI can compress these timelines. It can help us achieve the impossible in terms of understanding diseases that we've never understood before. But none of this happens overnight.

**[0:07:38] Luke Timmerman:** Okay, so the phrase cure all diseases in five years is almost completely wrong in every respect or not realistic.

**[0:07:46] Daphne Koller:** I wish it were true. I wish we could get there. I do think that in five to ten years we will be looking at diseases that today we consider to be impossible and have a path forward for a good fraction of them. I think that will be an incredible accomplishment. And cure all diseases in five years is just unfortunately physically impossible. You couldn't even go through the process of the clinical trials. There's not enough clinical trial sites.

**[0:08:15] Daphne Koller:** It's just, it's not physically a possibility to get there. So I wish people would remain calibrated and not give patients false hope. Because people who are sick are desperate to believe. And I think it's not fair to them to give them hope that is not achievable.

**[0:08:38] Luke Timmerman:** So you wrote a couple articles over the summer about there being no magic wands, which I think is taking on this perception that you see out there. What were you hoping to get across in those articles?

**[0:08:54] Daphne Koller:** I was hoping to present a balanced view that on the one side retains the incredible opportunity that we find ourselves in, where for the first time, I think we have line of sight to taking a discipline that has been one of serendipitous happenstance discovery, where somebody happens to have a good idea, have an interesting result in an experiment, which is the edge end of the thread that you pull on and suddenly something exciting emerges and all of a sudden we have a treatment for a disease and take that into an engineering discipline where there is a systematic methodical process by which we go from one disease to the other to the other and have a roadmap of how to get

**[0:09:45] Daphne Koller:** from the starting line to something that is more likely to work than not, which is nowhere close to what our industry currently is. It's, you know, our success rate of a treatment that goes into the clinic is stubbornly shy of 10%. And that's not even the starting line. The starting line is a lot earlier than that. And I think AI is absolutely the most ungating thing that will allow us to get there. That's on the optimistic side.

**[0:10:15] Daphne Koller:** But on the other side, there's still a lot of work to be done to deliver on that promise. And that work involves data collection of the right type, the development of AI methods that are not your out-of-the-box LLM, magically intuiting biological insights that people have not seen, and then miraculously turning those into a drug within a year or two. And I think that, you know, presenting that balance of what is possible, but also what is not possible is what I was trying to do in those essays.

**[0:10:49] Luke Timmerman:** Okay. So you've been at work here on Insitro for about eight years, I think. For those folks who are not familiar, what was the original idea here? What was the aspiration or the big opportunity that you saw? And then maybe how has it changed over time?

**[0:11:15] Daphne Koller:** So actually what I just articulated on the optimistic side was what led to the inception of Insitro. because I realized back in, you know, 2016 even, that we were at the kind of convergence point of two incredible revolutions. One is the ability to measure human and cell biology at unprecedented fidelity, quality and scale. And that revolution, while not as visible to the general audience as the AI revolution, because there's no chatbot, if you will, for CRISPR, but those technologies are incredibly powerful in allowing us to interrogate and measure biology and what has always been rate limiting in my opinion in really unlocking biological insight is the fact that we just never had enough data of the right kind and so

**[0:12:09] Daphne Koller:** I saw on the one hand the opportunity to generate and collect massive amounts of data and on the other side a tool that was becoming increasingly powerful and allowing us to see things in high content data that have never been seen and frankly could probably never be seen by a human eye and bringing them together to unlock new medicines for patients. And so that was that that that realization was the inception point of Insitro. And even back in the day, I realized that a lot of work needed to be done to put that into, to make that into a reality.

**[0:12:48] Daphne Koller:** At the beginning of the journey, I had friends, including very, very smart friends who'd been in this industry for a long time, who asked me, why are you going to all of this trouble of building a wet lab? I mean, you're not a wet lab scientist. You're an AI scientist. Surely it would be easier for you to just build a purely in silico company that just leveraged the data that already existed and derived insights from that.

**[0:13:14] Daphne Koller:** And my response was, the data that is out there is a diminishingly small fraction of the data that we need. It was not collected to be fit for purpose for building AI models. And that means that there is issues and artifacts and batch effects in there that are going to be impossible to retrospectively clean up. And so we need to create a data factory that would unlock the value that I think is there.

**[0:13:38] Luke Timmerman:** So this is not like a large language model in which you can just scrape off every bit of text, audio, and video that's ever been published on the internet since 1993 and put it into, feed it to an algorithm and train it over time and come out with a novel or a painting or something. This is not that kind of data problem. The data that you're talking about is not available. No one can just hoover it up and analyze it.

**[0:14:11] Daphne Koller:** No, it is not available. Data that is out there is often quite broken. As you know, science has a drastic reproducibility crisis, depending on whose estimates you believe. Somewhere between 50 to 80% of scientific literature is not reproducible. Some of that is downright, you know, whatever, misconduct. But a lot of it is just that science has a lot of dimensions and subtle things can make an experiment behave one way versus another when you're working with living systems.

**[0:14:40] Daphne Koller:** And so a lot of, and it's so reconciling all of that often incorrect literature is really difficult. And the higher up you get in terms of diseases that are uniquely human and are really complicated, Like anything, for example, in neuroscience, neurodegeneration and so on, doesn't even exist in other species. So the data has never even been collected in the right system at the right scale. And so there is no internet for us to rely on, to just hoover it up, hope that it's sufficient scale and quality, and then just go and connect the dots.

**[0:15:24] Daphne Koller:** I mean, AI is amazing at connecting dots, but if the dots don't exist, then everything is out of distribution.

**[0:15:33] Luke Timmerman:** So one of these early insights was you needed your own wet lab. You needed to generate data sets. You needed to create your own data, for starters, on what's causing disease. So how did you do that? How did you build this platform?

**[0:15:51] Daphne Koller:** So we have been building, and it's an ongoing build. Every time we keep making it better and adding to it and addressing, you know, making every piece of it, it's kind of like a constantly evolving engine where you continue to tweak and optimize the different pieces for printing biological data at scale. And that involves onboarding different cell types that speak to different disease biologies, perturbing them with the kinds of exposures that bring disease to the forefront because if you put cells in rich medium they are very happy and they don't exhibit any disease phenotypes and so they're not very useful so you need to stress them in the same way that they're stressed in our human body it's the stress

**[0:16:35] Daphne Koller:** that in some sense is is the oftentimes the inception of the disease and then you need to understand what are the causal factors that enhance the disease or revert the disease and And so we put in place also this genetic screening approach that allows us to perturb every gene in the genome and see the consequences of that perturbation in those cellular disease models. And the cells that you need to mimic disease are often ones that are harder to work with.

**[0:17:06] Daphne Koller:** So, for example, you cannot see neurodegenerative disease in a cancer cell. You need to have a motor neuron or a cortical neuron or microglia that resembles real human biology, and those cells are more finicky and harder to work with. And so there's a lot of different pieces that needed to be put in place in order to have the data that we collect from our cell factory really give us meaningful insight about disease biology.

**[0:17:36] Luke Timmerman:** And these are individual cells, which are useful in one respect, but it's not a whole organism. It's not in vivo data. Do you have some of that sort of data too?

**[0:17:46] Daphne Koller:** So that was the other realization. And frankly, one of the things that make us relatively unusual in this industry is the fact that we have brought together both the cellular data, which gives you a perturbable platform, which you need to do experiments. And on the other side, we bring in large amounts of human cohort data where we can see what happens in an actual living organism. And it's not just any old in vivo data. It's in vivo data in a human because, as I said, most diseases that we care about don't manifest in rodents.

**[0:18:17] Daphne Koller:** And so, you know, rodents are neither translatable nor sufficiently abundant. So they kind of have significant limitations as data collection systems. And so what we do is we collect large human cohorts. Now, the nice thing about those human cohorts, because you might ask, well, but you can't do experiments in a human being to understand what they do in response to a perturbation. I mean, you can do that experiment, but you do it at the very end and it's called a clinical trial and it is very highly regulated and rightfully so.

**[0:18:49] Daphne Koller:** But the good news is that each of us is an experiment of nature. Each of us has had our genome perturbed by nature at birth, and that has given rise to different clinical consequences of that genetic makeup. And when you look at it at a population scale, you can start to see, again, what is the genetic underpinning of having a higher or lower propensity to different kinds of disease biologies or greater protection against those disease biologies.

**[0:19:18] Daphne Koller:** And so what we've done is really built a two-sided platform with these two anchoring pillars. One is the cell biology where we get to do the interventions. One is the human clinical cohort where we get to understand what happens in an actual human being and the thread that ties them together. and also what gives us the ability to talk about cause and effect is changes in human genetics that give rise to changes in clinical consequences on the one hand and cellular consequence on the other and that is what allows us to bridge between what we see in our cell-based systems and what we

**[0:19:54] Luke Timmerman:** believe will translate to a whole human organism. So this is throwing off new hypotheses of potential drug targets, things that are going wrong and causing disease. So how many of these new target hypotheses have you been able to generate?

**[0:20:15] Daphne Koller:** So we have generated, you know, we have nine programs that are advancing into drug discovery. And frankly, at this point, and we're very excited about them, there's several in ALS. And I would love to have an opportunity later to talk about our ALS program because it's is such a clear demonstration of how AI can help you do the impossible, because as I'm sure you know, there's no disease-modifying treatment for the vast majority of ALS patients.

**[0:20:45] Daphne Koller:** Except for SOD1. Except for SOD1, and I saw recently a really interesting result with a gene therapy for FUS patients, an even rarer population, but those are the minority of cases where there is a very clearly defined genetic lesion and you're going directly after that genetic lesion to revert that. The vast majority of ALS patients are sporadic individuals for whom there is no clearly understood genetic basis and it's probably an interaction between some genetic factors and environmental stressors.

**[0:21:20] Daphne Koller:** And we have been, we believe, able to find not one but three programs that truly are disease-modifying in the 97 % of ALS patients that have what is called TDP-43 pathology, which is shared among the vast majority of both sporadic and most of the familial ALS patients as well.

**[0:21:46] Luke Timmerman:** I want to come back to the ALS program, like specific product candidates in a minute, but I still want to understand the underlying platform you've been building and modifying it over time. You've got all these technologies for single-cell analysis and what happens when you perturb one gene versus the next in a cell. And then, you know, how does that match up with observable human beings with the disease? Trying to bridge some of those gaps or connect those dots, if you will.

**[0:22:17] Luke Timmerman:** But how much data overall have you collected and where have you got it from besides the public domain? Have you gotten some from these large pharma partners that have all these, they have a bunch of that data too, right?

**[0:22:29] Daphne Koller:** Yeah. So we have at this point onboarded, I think, 28 petabytes of data. The majority of that is actually cell-based data because that is, you know, when you're tracking things, including imaging and transcriptomics at the single cell level, each cell spews out a lot of data. And so it accumulates quickly. But we also have somewhere between three and four of those petabytes are human clinical data. And some of those are from sort of semi-public resources like national biobanks, UK Biobank, all of us and such.

**[0:22:58] Daphne Koller:** Some of those are from specific partnerships with academic medical centers that have given us data that is specific to a particular disease biology. So one that has been disclosed is, for example, our partnership with the Moorfields Hospital in the UK that has one of the world's largest collections of OCT data, which is a form of eye scans. And it turns out that in those eye scans, you can see eye disease, but you can also see a lot of other diseases as well.

**[0:23:27] Daphne Koller:** It's not for nothing that people say that the eye is the window to the soul. It's also the window into your brain and, as it happens, other parts of your body as well. And so we are interrogating those OCT images together with our partners in Norfields Hospital to understand how disease biology manifests in those modalities. And so those are brought together. As I said, we don't print humans. And so the human data that we have comes from these partnerships.

**[0:23:55] Daphne Koller:** By and large, pharma doesn't share human data. There are restrictions about that. The data that we've gotten from pharma partners has often been of a different type. Like, for example, our really exciting partnership with Lilly that allowed us to predict in vivo pharmacokinetic properties for small molecules. But that's a sort of different prong of our work. Human data is not typically obtained from pharma partners. It's typically obtained from, as I said, either national biobanks that are specifically geared for enabling research and some partnerships with more specific providers.

**[0:24:32] Luke Timmerman:** Okay, so when you've got all this data, do you develop your own proprietary AI models to analyze it? Or are you using some of the foundation models out there commercially available?

**[0:24:45] Daphne Koller:** So the answer is yes to all of the above. So first of all, you know, AI is now a field where it's developed a set of building blocks. And obviously when we're building a new, for example, computer vision model for microscopy data, we don't start with a blank sheet of paper. we onboard the best segmentation algorithms and representation learning algorithms that have been developed for other imaging problems and adapt them to and expand them to deal with microscopy data.

**[0:25:10] Daphne Koller:** And often that adaptation is non-trivial, but it would be foolish for us to start with a blank sheet of paper. But at the same time, we've tried those models as they come out of the box, and they're not as good for microscopy data as they are for cats and dogs and airplanes. And so So we did have to do a fair bit of additional adaptation to get those to work well. And each of the data modalities that we work with, frankly, benefits from having special purpose models.

**[0:25:38] Daphne Koller:** You can take, you know, cardiac MRI and try and feed it into your favorite LLM. And it'll certainly say something, but the chances that that something is the best information you can get from a cardiac MRI, which is, by the way, a longitudinal time course, is not very high. And so you definitely get a lot more bang for the buck by developing specialized models for those specialized data sets. Having said that, on top of this all, there is absolutely, and not only on the top, we absolutely make use of LLM foundation models.

**[0:26:11] Daphne Koller:** We have agentic orchestration of a lot of our workflows that help us, you know, make sure that everything from low-level tasks, like, oh, here's our microscopy data. Can we rapidly QC it in order to see if it's sufficiently high quality? An agent turns out does that better even than a machine learning model. And so there's, you know, uses there. We use it to rapidly do experimental design and execution because it is able to rapidly sort of design, test, and then reiterate and so on.

**[0:26:46] Daphne Koller:** And that took a process that used to take our scientists like six to nine months and shrunk it down to, you know, eight weeks. And that's amazing. And then, of course, at the end of the day, when we're doing kind of like the holistic integration of all of the insights from all of our causal models, the agentic systems that are able to take all of that information, put it together, incorporate insights from the literature so that we're not completely oblivious to what has already been found out in biology.

**[0:27:17] Daphne Koller:** That is something that is the agents do that far better than a human mind can accommodate because there's just so much information there. So we absolutely have an incredible tailwind from those sophisticated LLM models.

**[0:27:30] Luke Timmerman:** Okay, so that's where some of the AI magic is benefiting you. But it's not necessarily like where the rubber meets the road and actually discovering new targets and figuring out like a small molecule or an algonucleotide to perturb that. That's still very much the hard work that

**[0:27:47] Daphne Koller:** you do day in day out it is and and there i mean if even if you look at some of the recent published work from the big frontier labs where they solve biological problems most of that if you look at how it was done is a an agentic system multi-agent system that runs tools that are specialized to specific data modalities that exist in biology it's not that they take all of the raw data put it into your favorite LLM and outcomes, the conclusion, these specialized models, all of which, by the way, are also AI models.

**[0:28:23] Daphne Koller:** It's just they're not like generic LLM. Those are absolutely critical for the, you know, the depth and quality of the insights.

**[0:28:33] Luke Timmerman:** Okay. But now you are on the cusp, I think, of bringing your first programs into clinical trials, the ultimate proving ground to test these hypotheses. whether you are right that these targets are indeed causing disease and that you can alter, you can modify that disease. Which is your lead program and how did you decide to work on that one?

**[0:29:01] Daphne Koller:** So we have three programs that are heading into the clinic in the next 18 months. The first one is our lead program, which is in MASH. It actually emerged from work that we did with Gilead, who was our very first partner, and they were very interested in liver disease in MASH, and we did some really exciting work together. And then they decided for, you know, big pharma strategic reasons that are often things that happen, that they are exiting the liver space and focusing on virology and oncology, and frankly, left us holding some really interesting, Not insights per se, those were not there yet, but a really incredible technology stack that allowed us to get to insights very quickly.

**[0:29:47] Daphne Koller:** And so the first target came out of that kind of foundational work that we did with Gilead. And there are actually three different MASH programs that we're currently pursuing, the first of which is the one that's heading into the clinic in very early 2027. So we're going to file this year and first patient in will be early in 27. And that came out of a combination of both the ability to interrogate human clinical data on the one hand and see liver fat and liver fibrosis at an unprecedented scale in large human cohorts that allowed us to do for the very first time a powered genetic analysis of those critical complementary traits.

**[0:30:32] Daphne Koller:** and identify a target that was a master regulator of both of those phenomena. And that target is the one that is going into the clinic. And then, of course, we were, sorry, I've neglected to say, that target also was then triangulated across, does it have the right effect in the cellular system that we have? So there was also a parallel screen for liver fat and liver fibrosis. And it's the triangulation of those two lines of evidence that helped us prioritize that target.

**[0:31:01] Daphne Koller:** And then, you know, I told you that I'm not a huge believer in animal models for discovery, but in this particular case, liver fat and liver fibrosis is a biology that is actually reasonably well conserved in rodents. And we were able to show a reversion of both the fat and fibrosis phenotype in translatable animal species. And so we have a lot of confidence around this particular program.

**[0:31:23] Luke Timmerman:** Well, there's no getting around it. You have to do the animal studies anyway.

**[0:31:26] Daphne Koller:** Well, you do for safety. You don't for efficacy, but it's expected if there is an animal model that is reasonably translatable, then people do expect it.

**[0:31:36] Luke Timmerman:** Okay, but you got more data from the rodents, and that looks to be a relatively predictive model as far as animal models go. And that bolstered your confidence to say this is a clinical candidate. Okay, so what's the name of the target?

**[0:31:50] Daphne Koller:** This is the only one that we've disclosed. So I'm going to answer this one, but if you ask me for the next ones, I'm not. So this one is IRS-1, and the genetics are incredibly strong for this.

**[0:32:05] Luke Timmerman:** And is this a small molecule or oligo or what kind of?

**[0:32:09] Daphne Koller:** A liver-targeted siRNA, which is actually really critical because you would not want to inhibit IRS-1 systemically. the consequences of that would not be great. But if you inhibit it in the liver, it's actually both very safe and very effective. Well, and as it turns out, siRNAs tend to concentrate in the liver. That's a feature. Yes, indeed. Especially when you add a galanac to them that targets them to the liver. And so they really don't go anywhere else at that point.

**[0:32:38] Daphne Koller:** And so it was a perfect match.

**[0:32:41] Luke Timmerman:** So you believe you've discovered this target? Do you have any competitors here?

**[0:32:47] Daphne Koller:** Not for this, not as far as we know. Obviously, people are very loathe to disclose targets of preclinical programs at this point, so I can't tell you that nobody else is working on it, but we don't know of anybody.

**[0:32:59] Luke Timmerman:** Yeah, well, why did you decide to disclose it if there's reasons to keep it quiet?

**[0:33:04] Daphne Koller:** You know, I think at the time that we published the paper, so we published a paper, and we, sorry, actually, let me reframe that. We filed a patent in order to protect the composition of matter, And for an oligo therapeutic, the sequence that you are disclosing in the patent tells you the target. And so when the patent published, it didn't make sense to not disclose the target because anyone could have inferred it anyway.

**[0:33:33] Luke Timmerman:** Okay. And you need the patent for obvious reasons. If people are going to invest in this over time, you have to have some belief that you'll be able to earn income from it if you're successful.

**[0:33:45] Daphne Koller:** Yeah, that's exactly right.

**[0:34:14] Luke Timmerman:** universe. That's clinical trial data, peer-reviewed literature, regulatory filings, conference presentations, and expert transcripts with researchers, clinicians, and executives. And it keeps monitoring your space even after you've closed your laptop. It's designed to act like an extension of your research team. It remembers the programs and competitors you're tracking, follows your process, and applies it automatically to new data as it comes in. And every output links back to its original source, so when you're citing a trial result, you can see exactly where the number came from.

**[0:34:46] Luke Timmerman:** Take the task you've been meaning to get to, tracking a competitor's trial program or flagging the next regulatory decision before it hits, and hand it to superanalyst. Start your free trial at alpha-sense.com slash the long run. All one word. That's alpha-sense.com slash the long run. And quick question, when was the last time a CRO told you what something costs before you sign an NDA, sat through two scoping calls, and waited a week for a scope of work?

**[0:35:19] Luke Timmerman:** In bioanalysis, pricing gets treated like a closely guarded secret. Dash Bio thinks that's backward. It publishes pricing publicly on its website and even has a calculator that shows how the price changes if you redesign your study. You can see what your study costs before you talk to a human being. No surprise line items, no it depends, no games. Dash was built to be the bioanalysis company that's actually straightforward to work with. Transparent pricing, guaranteed timelines, and data in days, not months.

**[0:35:50] Luke Timmerman:** If you're tired of pricing that feels like a negotiation every single time, run your bioanalysis with Dash. Visit dash.bio slash pricing and see exactly how much your next bioanalysis study will cost. What kind of magnitude of benefit do you think this could provide for patients with MASH, which is metabolic-associated steatohepatitis? I think I got that right. Yes, you did. A fatty liver disease, inflammation, fibrosis of the liver. Millions of people have it. There's maybe one, I think, FDA-approved product at the moment.

**[0:36:29] Daphne Koller:** Yes. You know, people don't talk very much about this disease because it's silent until it basically comes close to killing you. And so I think it's not as obviously visible to people in the same way that, say, ALS or other diseases might be. But the unmet need is quite large. Incretins, for a while, there was some presumption. And when we talked to investors, like, yeah, you know, incretins are just, you know, GLP-1s are just going to cure it all.

**[0:36:58] Daphne Koller:** And it turns out that, yes, there is an effect of the GLP-1s on liver fat, but that effect wanes off relatively quickly. And the amount of liver fat reduction is not dramatic and certainly not sufficient for a very large number of individuals. Not to mention that there is a fairly significant fraction of individuals who have what's called lean mash, which basically are people where their liver tends to accumulate fat regardless of whether they are overweight or not.

**[0:37:25] Daphne Koller:** And so the incretins are actually not a very good intervention for them. And so we believe that there is still a significant role for this therapeutic, both in maybe as a combination for the individuals where there isn't like an excess weight component to the disease, but the reduction of the liver fat that's needed is more profound. And for the fairly significant fraction of lean-mash individuals that incretin wouldn't be helpful for. Now, the other thing I would say is that we believe that this is a very well-tolerated therapeutic, and a lot of the drugs that are currently under development have dose-limiting side effects that make it less appealing for patients.

**[0:38:13] Daphne Koller:** And so there's that as well. And finally, maybe less exciting but still important from a patient experience perspective. SIRNAs have very long durability. This one is easily three to six months injection, and that's much friendlier to patients than having to inject yourself once a week.

**[0:38:31] Luke Timmerman:** Yeah, there are a lot of reasons why siRNAs are a pretty compelling modality. You outlined them. So you worked on this program with Gilead for a while. They decided they weren't going to do it. So you own this one 100 %?

**[0:38:45] Daphne Koller:** Yeah, we didn't really work on this program. We worked on a lot of foundational work for interrogating MASH and had gotten just to the point where we were going to start, you know, nominating targets and exploring them when Gilead basically decided that this was not a priority for them. So this is, Gilead has no rights on this. The target didn't even exist at the time that the partnership was in action.

**[0:39:12] Luke Timmerman:** Okay, so you owe this one 100 % and are taking that forward into the clinic. Yes. But you mentioned earlier ALS, amyotrophic lateral sclerosis, this progressive neurodegenerative disease that is terrible. And many companies have tried and failed in this area. there are a couple of rare genetic forms of it that people have been working on and have had some positive results. But for the vast majority, there's nothing. What did you find here that others have not?

**[0:39:52] Daphne Koller:** So the majority of ALS patients have something called TDP43 pathology, which is a particular protein called TDP43 that has a role in the nucleus. Specifically, it modulates splicing of RNAs. And the hallmark of the disease is that it's exported out of the nucleus. It segregates in the cytoplasm, and it doesn't execute its function anymore. And so in ALS patients who have TDP-43 pathology, basically a lot of defects occur in the disease in which you have a lot of misplicing events because TDP-43 that's supposed to be doing a really critical function is not doing that function anymore.

**[0:40:39] Daphne Koller:** And that has consequences both in, you know, there's detritus of, you know, protein fragments that shouldn't be there that are there and are clogging up the machinery. So these are called toxic cryptic exons. And then on the other side, you have full length versions of those same proteins that are supposed to be doing a really critical function in your own health and are not doing that function. And so that is a hallmark of 97% of ALS patients.

**[0:41:07] Daphne Koller:** We recreated that disease state in our motor neurons after confirming that, in fact, you did get that disease state in motor neurons from ALS patients with different genetic backgrounds. And so we confirmed that this is, in fact, the hallmark of the disease. And we were able, by seeing these motor neurons in our high-throughput cell factory, identify disease access. What does ALS look like in a cell? What does a disease cell look like relative to a healthy cell?

**[0:41:39] Daphne Koller:** And then identify a set of reversions, a set of genetic interventions that cause the neuron to look more like a healthy neuron than a disease neuron. And now you could say, well, great, so it looks healthier, but is it really healthier? Which exactly is the question that we and our colleagues at BMS asked. And we confirmed that you have a significant reduction in the cryptic exon, those toxic cryptic exon, and a significant increase across the board in the full-length versions of the proteins that are necessary for neuronal health and disappear in these ALS neurons.

**[0:42:23] Daphne Koller:** And so what we found is basically a master regulator that buffers the effect of this TDP43 pathology, which exists in all of these patients. Because people tried to go after TDP43 itself. But it turns out that some of the interventions that people tried were, well, let's clear out TDP43. Well, when you think about what I just explained, that actually is going in the wrong direction. That is, you can clear out the TDP43 aggregates, but it doesn't restore TDP43 function to the nucleus.

**[0:42:52] Daphne Koller:** So it's not actually helping. And people tried to dissolve the aggregates, but that didn't cause TDP-43 to go back to the nucleus either. And so what we did was we basically found a buffering mechanism that allows these ALS neurons, even when they have TDP-43 pathology, to function far more like healthy neurons. And the really definitive experiment that is like that kind of like wow moment was when we did a really finicky and complicated experiment for functional recovery.

**[0:43:26] Daphne Koller:** The hallmark of ALS is that, you know, when these neurons become sick, those neurites that connect the neurons to muscle start retracting and they don't grow back. And they basically detach from the muscle and are stopping to control the muscle function. and so we did an experiment where we sliced off those neurites and asked do they grow back and in healthy motor neurons they do and in ALS neurons with TDP43 pathology they don't and with our drug they do grow back and that to us was like oh my god we have recovered the fundamental functional deficit of a diseased motor neuron and that is what gives us this deep deep optimism that we have found a disease-modifying therapeutic.

**[0:44:14] Daphne Koller:** And if I can just add one note, because this is not our contribution, but it made me so optimistic. I read some of the reports around the Topher syndrome that was provisionally approved. This is a biogen drug specifically against SOD1 patients, which you alluded to earlier. And there are reports that it doesn't just slow or stop the progression of the disease in a lot of patients. It actually reverts the disease in some patients. And you would ask, well, how is that?

**[0:44:45] Daphne Koller:** I mean, if the motor neuron's dead, it's dead. Well, it turns out that when motor neurons die, it takes them a while to die. And if you're able to intervene early enough, you can bring some of those back to a healthier state. And so people recover some function. I mean, you still can't bring back the dead motor neurons. But if you intervene early, then you can make some of the ones that were kind of like teetering on the brink and pull them back.

**[0:45:10] Daphne Koller:** And I think that is just such an exciting thing if we are able to do that with our drug. And I really hope that we'll be able to show that as well.

**[0:45:18] Luke Timmerman:** Well, these whole neurological systems are just so adaptable and complex. I mean, we have neuroplasticity in some respects. And so, I mean, if you're able to restore some or kind of create a workaround around the dead neurons so that the function is taken over by another one or you can regrow other ones that are still relatively healthy. I mean, gosh, a lot of things become possible.

**[0:45:42] Daphne Koller:** Yeah, I'm not sufficiently in neuroscience to be able to say how much of that is actually possible and so on. But as I said, the reports that I've seen on functional recovery for some patients really give me optimism about the potential of any intervention that actually brings sick neurons back to health. And since that is exactly what we screened on, is bringing sick neurons back to health, I'm hopeful that we'll be able to do that.

**[0:46:17] Luke Timmerman:** Okay. And again, this is an oligonucleotide?

**[0:46:20] Daphne Koller:** Yes. This one is what's called an antisense oligo versus an siRNA. And the reason is that that's the modality that has been most de-risked for neurological interventions. There are some approved drugs that use an ASO, and we made the decision to not explore new modalities in this context because we really wanted to focus on, you know, being as de-risked as possible in the therapeutic modality and really targeting the biology.

**[0:46:52] Luke Timmerman:** It's a single-strand gene silencer versus a double-strand gene silencer. And, you know, people have their preferences. But, yeah, there's a reason why you might proceed here. And now, is this the one that's partnered with Bristol-Myers Squibb?

**[0:47:07] Daphne Koller:** So, yes. And this is a little bit complicated. So, what happened was three targets emerged from that partnership. It was an incredibly productive partnership. Oftentimes in these, you know, pharma biotech partnerships, you know, there is a discovery agreement and nothing comes out. Here, three targets were nominated. And BMS elected to pursue the first of those targets, ALS-1, as a small molecule. Because I think they were, you know, concerned about an intrathecal oligo from a commercial perspective.

**[0:47:40] Daphne Koller:** And we came and said, look, this is a really hard small molecule target. We think it's doable. We actually made a lot of progress towards having a small molecule therapeutic against that. But it's going to take a while. And we said, there are people dying. And so how about you let us, at our cost, on our dime, take that same target and advance it into the clinic with an intrathecal oligo. And we were, I think they too share our deep commitment to trying to reach patients as expeditiously as possible, especially in the context of such a grievous disease.

**[0:48:17] Daphne Koller:** And so we're very grateful that they let us do that. And that enabled us to advance this program really, really quickly. And we are on the cusp of declaring a development candidate for that program. And then, of course, you need to do all the whatever IND enabling studies to ensure all the GLP safety and such. And so it'll take a little bit of time after that. But then we hope to get that into the clinic in early 2028.

**[0:48:46] Luke Timmerman:** And you really think that this could be a treatment for 97 % of ALS patients?

**[0:48:52] Daphne Koller:** Well, that is the hope. Like I said, 97 % of ALS patients have TDP-43 pathology. Look, I don't want to tell you it's going to—I mean, I remember we discussed at the very beginning of this episode this notion of the word cure. I am very reluctant to use the word cure because I think it holds out so much. There's so much baggage associated with that and such an incredible set of expectations. But will it change the course of the disease for a very large fraction of ALS patients?

**[0:49:22] Daphne Koller:** That is certainly the premise that we're working under.

**[0:49:25] Luke Timmerman:** Uh-huh. Now, I know you said you had nine programs, and I don't want to talk about all of them, but there's another one that you're working on with brown fat. What is that one, and how did that rise to the top of your priority list?

**[0:49:38] Daphne Koller:** Yeah, so that one, I think, was the brainchild of Dave Lloyd, who's our SVP of metabolic disease, and he's a brilliant scientist and has a track record of, you know, really interesting programs earlier in his career. And I think he has always had the hypothesis, which now I think many others are sharing, that while the incretins are amazing, and by the way, he worked when he was at Amgen on Meritide, which is their incretin, their GLP entrant, which is apparently doing really well in the clinical trials.

**[0:50:16] Daphne Koller:** but he understood that it's one thing to shut off the the supply of calories by having people eat less but even but when you look at people there are people who eat the same amount and some of them stay really skinny and others you know as much as look at a jelly donut and you know gain 15 pounds so why is it that even when you correct for the same caloric consumption Some people gain weight much more easily than others.

**[0:50:44] Daphne Koller:** And some of that has to do with how effectively they burn calories. And that would be a very important complementary therapeutic to the incretins. And one is likely more muscle sparing as well, and also better tolerated from a lifestyle perspective. And so we started to work on peripheral obesity mechanisms that focus more on how calories are consumed rather than how many are consumed. And so people have had for quite a while this hypothesis around brown fat, which is something that exists in babies with very considerable amounts.

**[0:51:24] Daphne Koller:** And then, unfortunately, we all lose it as we age. Brown fat is a metabolically active form of fat. It burns calories rather than stores them. and in rodents it's been shown that the more brown fat the mice have the healthier they are not only from a weight perspective but also from multiple other aspects of cardiometabolic health liver function kidney function um you know multiple other cardiovascular health and so on type 2 diabetes but we were never able to measure it in humans at scale and so people hypothesized that it also played a role in human beings, but we weren't able to necessarily show that.

**[0:52:04] Daphne Koller:** And so what we were able to do using our AI stack was basically see brown fat at scale for the very first time from whole body MRI and measure it. And cool things emerged. I think the wow moment for us there was when we looked, one of our scientists, our data scientists had the idea of looking at the amount of brown fat that people have in MRI relative to the time that the MRI was taken. And it turns out that if the MRI was collected in winter, you have more brown fat.

**[0:52:35] Daphne Koller:** And if it's collected in the summer, you have less brown fat. And so all those fitness people doing cold plunges in order to stimulate their brown fat, they're onto something, except that you'd kind of probably have to live in the ice bath. And that's not very fun. But it turns out that, yes, you do, in fact, have brown fat stimulated by cold. And so, but with that insight, we were now able to go back and remember genetics is our north star for causality.

**[0:53:01] Daphne Koller:** And so we were able to go back and for the first time identify the genetics of why some people have more brown fat than other people. And that gave us a really interesting set of targets, several of which are now under investigation. But one that stood out to us as being particularly interesting, we put into a mouse model and And we're able to show that that mouse lost 25 % of its fat mass. It didn't lose any lean mass.

**[0:53:30] Daphne Koller:** And it didn't stop eating. So it did not operate through caloric restriction. And so that gave us a lot of excitement around that mechanism and caused us to continue to advance that program. although as I said at the same time given that we were able to do the first ever power genetic analysis of brown fat there's a slew of other targets that we also have excitement around that we're now interrogating both in cell systems and in animals.

**[0:53:59] Daphne Koller:** Now how far away

**[0:54:01] Luke Timmerman:** is one of these programs from entering the clinic?

**[0:54:04] Daphne Koller:** Those are a little bit further out I don't want to give you a timeline it's not going to be I think, you know, early 2028. The one that is in the same timeline as the other programs that we spoke to earlier is actually one in dry AMD, which is age-related macular degeneration. It's the largest cause of blindness in adults over 50. Really horrible disease. And you talk to a lot lot of people and many will tell you that you know they getting going blind is the scariest thing that they could possibly imagine and we were similar and this was something that emerged both from that earlier partnership that we spoke about which is the one with more fields hospital

**[0:54:57] Daphne Koller:** that we were able to train the world's largest foundation model for oct data and use it to understand sort of the underlying biologies that exist in dry AMD, and that's what gave us this target. And that one is also on the cusp of a development candidate at the end of this year, and will also go into the clinic in early 2028 once we complete the IND enabling studies.

**[0:55:21] Luke Timmerman:** Oh, the dry form of age-related macular degeneration, not the wet form for which we already have VEGF inhibitors that have been around and are quite successful, help people see more lines on the eye chart.

**[0:55:32] Daphne Koller:** No, absolutely, yes. And that is, although I will say that for a subset of patients, when you address the causes of the wet AMD via VEGF, which is this excess vascularization, they develop dry AMD. So we still need that intervention even for a subset of those patients. But yes, dry AMD, the current set of treatments are only modestly slow of a disease progression.

**[0:56:00] Luke Timmerman:** Okay. Now, let's zoom just back out a little bit at the bigger picture here with AI for drug discovery. You know, you follow a lot of the other companies, I'm sure. You see them at investor conferences, both ones who've been around 10 years and others that are newer. were, what do you see happening in the industry that gives you optimism? That at least among the people who are really grounded in disease biology, that they're on to something and they're going to get somewhere positive in the next, say, five to 10 years.

**[0:56:33] Daphne Koller:** So I think the thing that gives me the most optimism actually is not the work that's most grounded in disease biology, but the work that has focused on the next stage of the process, which is also really important, which is once you've identified a mechanism that you'd like to go after, how do you turn it into a drug? And, you know, for an siRNA, it's a relatively straightforward linear process. But when you start getting into large molecules and small molecules, these things always took time and in many cases also had an uncertain outcome.

**[0:57:08] Daphne Koller:** And there's a slew of companies that are now really creating a path, a technological path to go from a target to a drug very, very quickly and effectively. and that is really valuable because all of us once we've identified what we want to go after would like to get there as quickly as possible and with as few side effects and with the right pk properties and so on and so forth that allow something to be not just a binder but actually safe and effective in a human and so i think that is an area where i see tremendous amount of legitimate uh exuberance around the progress that's that's happening coming back to the beginning of our conversation, I think people

**[0:57:52] Daphne Koller:** often lose sight of the fact that that is only one part of what is typically viewed as a three-part process, which is first you have to have the right target before you make the drug. And then you have to, once you have the drug, you still have to execute the clinical trial in the right way. And both the first part and the third part really do require a much deeper understanding of disease biology and human pathophysiology.

**[0:58:15] Daphne Koller:** and I see less effort and progress on that in many companies. But the middle part is truly being revolutionized, and I'm very grateful for that.

**[0:58:26] Luke Timmerman:** Now, is there a new balance to be struck between, say, open source biology and then what happens behind proprietary walls at companies like yours? I mean, you sometimes see these announcements. I mean, there was just one about the biohub recently to do cellular perturbation data and put that out in there in the open. What's your view on the right way to balance open versus proprietary here to advance the field?

**[0:58:53] Daphne Koller:** So I think that it's really important that, you know, both governments and these philanthropic organizations make an effort to create large resources that enable research to be performed, especially among organizations that are less well endowed financially, but have really, really smart people such as academic institutions and smaller companies. So this type of cellular resource is one important example, and I applaud the effort. The other that I think is arguably even more critical is creating additional and richer human cohort data resources, such as the UK Biobank, All of Us, and so on and so forth.

**[0:59:40] Daphne Koller:** These are few and precious. And human data is one of those things that is really impossible for even a single company, even a very well-endowed pharma company to collect. There's just not that access to patients and especially longitudinally tracked patients. And so you really need to have a different type of organization that creates those resources of human clinical data appropriately consented, appropriately in a research environment that doesn't allow them to be downloaded willy nilly. All of that is obviously important.

**[1:00:16] Daphne Koller:** But that's the kind of data that we as a community are most lacking in and one where I wish there was more of an investment from governments and philanthropic organizations to create those. Because at the end of the day, Luke, the only real model system for a human being is a human being. And we just don't have enough data on human beings.

**[1:00:38] Luke Timmerman:** Well, as I was getting ready for this conversation, I started thinking about the nature of the data problem. How much data is there in the universe about human biology? I mean, it's astounding. I mean, there's something like 36 trillion cells in the average adult male, maybe 28 trillion, and the average adult female is a little smaller. We're producing 200 to 300 billion new blood cells per day. That's each and every person, 8 billion people on Earth.

**[1:01:06] Luke Timmerman:** We coexist with 30 trillion bacterial cells. and they're always like growing up and dying and turning over and having mutations, somatic mutations. I mean, the amount of data that is out there and the, and versus the amount that we collect. I mean, I think it would be generous to call it like a thimble full in the universe

**[1:01:26] Daphne Koller:** is what we have, what we have, not, not even. No, I completely agree. And again, coming back to some of the data disparities, the challenge that we have on that type of data is that the data that is most relevant to humans is the one that is the hardest to collect because it has to be collected in humans. So in the essay that you alluded to earlier, I talk about some of the differences to, for example, the data that informed alpha fold.

**[1:01:56] Daphne Koller:** Protein folding is actually closer to physics than it is to biology. And it works in pretty much the same way if you collect the protein in a bacteria, in a yeast, in a mouse, or in a human. And so you're able to bring in all of the data that we've collected around protein sequence and structure and use it to train the models. And it's a relatively simple model because, I mean, not to diminish from the incredible accomplishment there, but it's a static viewpoint of a single molecular entity at a single point in time.

**[1:02:26] Daphne Koller:** Human beings, as you said, are so much more complicated and they're dynamic and they're evolving and they depend on so many aspects of their environment and genetics and so on and so forth and the as i said and the processes that matter the most to human disease are the ones that you can only collect in a human you can't collect it in a bacteria or even in a rodent and so the data gap is just staggering but at the same time i think we shouldn't let that defeat our motivation because even with resources like the UK Biobank, which came nowhere close to every single cell and every single, you know, at the molecular level and so on and so forth,

**[1:03:08] Daphne Koller:** but just collected human clinical phenotypes with human genetics and some blood biomarkers and so on and so forth. The scope of biomedical insights that came out of that is staggering because we've never measured any of those things before in 500 ,000 individuals. So yes, the ceiling is very, very high. It's enormously high. But where we are is so low that any progress that we make unlocks tremendous opportunities for new medicines. And so I really hope someone just buys into that vision and unlocks that additional set of value from a richer data resource.

**[1:03:52] Luke Timmerman:** We're just scratching the surface, and yet there's so much to learn. Daphne Koller, thank you so much for joining me on The Long Run.

**[1:03:58] Daphne Koller:** Thank you so much, Luke. It's a pleasure, as always.

**[1:04:04] Luke Timmerman:** Thanks for listening to The Long Run, a production of Timberman Report. Pedro Rosado of Headstepper Media was a sound editor. Music is from D.A. Wallach. See you next episode. you
