David Fajgenbaum on drug repurposing, AI, and saving his own life

21:57I can't do anything. I'm sick and I'm never going to be better. So the idea that you had been read your last rites five times, we're like, I got this. I'm very much in awe. So that is incredible. And then what was that moment that every cure came into being? And how is that sort of changing the game at scale? Absolutely. So after this drug started working for me, as I mentioned, I just started thinking, oh, my gosh, how many more drugs are there out there? Then we started using this drug, serolimus, and other Castleman's patients. It was helping them. Unfortunately, it works in about 20 % of patients.

22:27So it doesn't work for everyone. But it actually worked in the first four patients, so me and then the next three. So I thought, I was like, oh, my God, we figured this thing out. It turns out we didn't figure it out fully. But that was important. And then we found another drug for Castleman's for a patient named Kyla, this young girl in Chicago who's dying from her disease. And nothing was working. But we tried another drug, and it worked for her. And I was like, oh, my gosh, we can actually find other drugs for this disease. And the next thing was that we actually found a drug for a patient with a horrible cancer called angiosarcoma.

22:55So now it was a proof point that, wait a minute, we can actually find other drugs for other diseases. And as time was going on and as we got to the early 2020s, my co-founder of Every Cure, Grant Mitchell, and I, Grant's a dear friend. I think you met him at a Do Something event years ago. He has the record for rock skipping. I don't know if you remember if he shared that story with you. He's a world champion rock skipper. but Grant is also a brilliant AI mind and he did his MDMEA with me and then and then went on to work at quantum black and doing a lot of great work in AI so Grant and I kept talking about okay David you're doing all this cool stuff in the lab at this stage we'd almost we were up to almost 10 different drugs that we'd repurpose for different diseases in my lab at Penn he said why don't we try to really scale this and automate it with artificial intelligence what if we can actually replicate what you guys are doing in the lab but actually do it at scale across all drugs and diseases.

23:46And so we kept talking about it. We were super excited about it. We were making good but slow progress in the few diseases we're working on. And then we spent time with amazing AI experts talking about ways that we could really scale this. And so in September of 2022, so three years ago, we created this nonprofit, Every Cure, to really scale this. And it's on a mission to save lives with repurposed drugs. And we do that by using AI to find new uses for every medicine we have. I frequently wonder about the naming of diseases because we kind of do it on a rough kind of set of symptoms to disease progression.

24:22Does the fact that you've actually discovered different drugs that work for different people with Castleman's actually mean there's multiple Castleman's diseases? And then it actually, the fact calling it all Castleman's is in a sense too macro? I completely agree with you, yes. Because if you treat a disease with an IL-6 blocker and it makes this patient better, then you know that that's like the underlying driver. If you're this kid with an mTOR and they get better, then – so to your point, like the molecular mechanisms are different. And so, yes, I think that we absolutely need to get to a world where we're actually subtyping and naming things based on like what the actual mechanism is as opposed to – you're right.

24:59Right now we name things based on a collection of symptoms. So, yes, I think we need to go there. What's challenging is that with this sort of collection of symptoms, we've got 18 ,000 diseases. I think if we were to get even further, it would be like, I don't know, hundreds of thousands of diseases. But I think that's probably the kind of granularity you need to really be able to find more treatments. Well, because, like, what should the additional blood test panel be? Exactly. Right. It should be based on what might be the treatment. Yeah. So tell us about your – I think your AI program is called The Matrix.

25:28Right. So tell us about it. What knowledge does it ingest? What does it accelerate? Sure. So what we really set out to do when we created EveryCure is that we want to unlock the full potential of every drug to treat every disease in every patient possible. That's our broader vision. And the way we're doing that is through this AI platform matrix where we are quantifying the likelihood of every FDA-approved drug, all 4 ,000, to treat every human disease, all 18 ,000. So if you were to create a matrix with drugs and diseases, there would be 75 million possibilities, 4 ,000 drugs times 18 ,000 diseases.

26:03And so we basically want to be able to help to prioritize among the 75 million possibilities what are the best ones for humans to move forward to help people. You know, not just what's the best drug for Castleman's or what's the best disease for this drug, but really what are the best opportunities to help people. And so we've created this matrix, all drugs versus all diseases. And our underlying infrastructure is we utilize primarily biomedical knowledge graphs, which, as you know, well read, map out the world's understanding of human biology. So our most recent knowledge graph is tens of millions of biomedical concepts as nodes, tens of millions of edges connecting those biomedical concepts.

26:38And basically, and I think I'm actually going to show you guys a demo a little bit later on. Basically, you can imagine mapping out all of that onto one two-dimensional space, every drug, every disease, every gene, every protein, and all the connections between all of them. This serves as our understanding of human biology. And within this understanding of human biology, we then train algorithms on known treatments. So we know that siltuximab treats Castleman disease, and we know that sildenafil can treat pulmonary hypertension. So we train on the known connections, and then we ask the algorithm to make a prediction on every other drug for every other disease and give us a score from zero to one.

27:13So if it looks really good, give it a 0.99. If it looks really bad, give it a 0.00. Yes. We would love to see a high score in action if you would mind pulling out your laptop. Yeah, let's do it. So now I've got my computer, let me show you a little bit. And we've got this up here on the screen. So two things. So first off, we have something called Orchard. We call it Orchard because we're looking for low-hanging fruit across the forest. And we call these treatments pears. It's like a drug-disease pear. So that's our Orchard. And so in Orchard, our tech team basically generates a score for every drug versus every disease, presents it to our medical team.

27:50And this is our view. So basically this is the top scoring drug for every drug versus every disease is this penicillin for sepsis, which makes total sense. Like this drug treats sepsis. Like this is what you'd expect. As you can see, one of our colleagues called it a known entity. Like that's what you want to see at the top of your list. What's really cool is that we have all these things at the top that you'd expect to see that score 0.99s. But then we have this what's called suggestions mode, which is basically for our medical team to review through and see things that are not obvious but are interesting.

28:18that the algorithm is giving a 0.99, but that it's not already a treatment for that disease. Then they go through and they say, I wonder why lidocaine might be scoring highly for breast cancer, you know, and then they dig into it. And one of the ways they dig into it is through these biomedical knowledge graphs. And so what I've got here, and I pulled up a few. So this is just a tiny little segment. I mentioned our actual knowledge graph has tens of millions of nose and edges, but I pulled up just a little tiny segment actually of breast cancer to lidocaine. This is breast cancer over here on the left, lidocaine on the right, and all of the connections that could explain why it might be a potential treatment for breast cancer.

28:56And so, you know, lidocaine is the numbing medicine you get when you go to the dentist. There's actually really interesting data that it has direct cancer-killing properties, and it happens to be the substance that we inject all over our bodies for every surgery all the time. So it's a quite inert substance. It has cancer-killing properties. And also there's even some really interesting data that it could prevent recurrence of breast cancer if injected around a breast cancer tumor before surgery. So for us, that's really exciting because this is a cheap old drug that's available already in the OR.

29:21It's already going to be used in the OR. The change here would just be to inject a bit more of lidocaine but around the tumor. So the question is, okay, well, how might it work? And the thought here is that through apoptosis, which is cell killing, connections between lidocaine, the genes and proteins that it interacts with, and then also the role that it might have in breast cancer. And so you can imagine we have knowledge graph or this is a sub graph for every connection that you could think about that our medical team can then use. And so the AI is, again, running through all these millions and millions of combinations.

29:51There are so many people who are who are in need of you. Like there's way more to do than you guys possibly have time. Are you triaging based on the number of people that have a disease just based on the top score? Like how do you think about what to focus on from a research perspective? Great question. This is the biggest challenge with the work that we're doing, and it's because there are so many options and there's so much potential. It's not like we have to find the one treatment out of 70 ,5 million. There's actually a lot in here. And so it's a big challenge. So the first thing is that the score that we get through the matrix from our biomedical knowledge graph and machine learning approach gives a score from zero to one.

30:27So the 0.99 is obviously we're most excited about, but all the way through low scoring. That's how likely is that drug that it will work in that disease. That's one input. But then to your point, impact is really important to us. We want to go after the really bad diseases because we've got limited time, limited resources. You want to go after the bad stuff first. So we also have generated something called an unmet medical need score. And so that unmet medical need score gives us an idea for just how bad the disease, things like ALS and pancreatic cancers, get the worst. And then things like toenail fungus get the least bad score, right?

31:00Sure. But it's linear. So we can really focus on the things in the top quartile. So we look for really promising biomedical impact score. We look for a really high unmet need score. And that's when our med team starts really digging into these things. And so if you look here at Orchard, so this is actually our live platform. And so you go into triage. So there's 204 things that our medical team has looked at and said, this looks really, really interesting. And I wonder why this thing is scoring highly. 62 have gone through what's called a med review where they've dug deeper into like this looks promising.

31:33Why might it work? 17 have had deep dives. There's 11 that have been endorsed by our advisory council. And so once it gets into the medical team's review, and we have this amazing team in Boston of MDs, PhDs, MD-PhDs, who review through these things. Once it gets to their review, this is where it looks promising enough. AI is pointing us towards it. There's enough unmet medical need that now we need to focus on. Is it really going to help people, and how do we get it to patients? Obviously, it's an amazing use of AI technology. But one of the things that people frequently overlook in these, it's AI plus human.

32:08Yes. So say a little bit about what the medical team does, how it's a human amplification in terms of a journey of discovery. Yes. I'm so glad you brought that up because we have humans in the loop in so many aspects of this. So our tech team, we've got a 20-member tech team that embedded within the tech team are a few MDs, a few PhDs who really come at it from a biomedical science perspective, less from a tech side. So they're embedded within the 20-member team. And so even when they're generating predictions or improving our models, they're thinking about this from a physician's perspective or from a researcher's perspective and thinking about, you know, what should this model look like to make sure we're really finding treatments that will work as opposed to maybe just biomedical concepts, right?

32:49So it starts very early on with the tech team having that sort of a lens. It also, to your point, once it gets to the med team, these are medical scientists and physicians who are really just thinking about it from a human perspective. Like, okay, it's great. We love when the algorithm goes at a 0.99. But now we've got to really think about, like, what are we going to do with this thing? Is there a clinical trial needed? Is there laboratory research needed? In some cases, we actually come across drugs that look really promising. And there's actually already been a lot of work done. It's just about getting the word out about that particular drug.

33:20And so our clinicians and our researchers on the medical team are really thinking about this from the end goal of reaching as many patients as possible with this condition. What's the kind of impact we would have if we move this forward? And part of this, there's multiple places where the human will look, decisioning, context awareness, being able to think about, okay, which of these things is potentially a really interesting genius idea and which of them is kind of actually just a nutty idea. because he's just kind of sorting through those. Can you do FDA filings yet using the fact that you've got kind of like a deep knowledge base, which is suggesting this potential off-label use?

33:59What does that journey look like? Because I presume that when you add an FDA-approved use, that gets magnified through the whole system. Absolutely. Well, maybe I'll start back at your comment around sometimes there's nonsense because not our platform, but actually another machine learning model that I've seen applied for drug repurposing. We were just looking through the top hits for Castleman's, for example. And I remember seeing that the fifth highest scoring compound for treating Castleman disease was car exhaust fumes. And I was like, okay, that's interesting. Maybe it's doing something with cytokines, but we probably shouldn't inhale car exhaust fumes, right?

34:36That's probably not a good way to treat it. And so to your point, there was something in the data that suggested that maybe IL-6 levels or mTOR goes up or down when you inhale car exhaust fumes, but we don't recommend that. And so I think that it's really important for these humans to review through these things and skip over that. But then to your question around, okay, you find something that looks really promising, how do you make sure it reaches patients? Yes, working with FDA to get the label changed so that it now, so that drug actually has that condition on the label, that's amazing. And that's really, really powerful and really important.

35:05And we are working with FDA and we're eager to continue to work with them to find more of these cases where drugs that look promising can actually get a label change. But even when we do work with them, I still think it's important to know that we would never go straight from AI prediction to let's give it to patients, let's get on the label. It's AI prediction, all right, let's confirm it in the lab. Let's make sure that the biological mechanism makes sense. Let's give it to patients in a well-controlled and appropriate an environment to prove that it works. And then let's talk about what the label changes look like.

35:38We're talking about this sort of AI plus humans to make this happen. And obviously, all of the frontier labs right now are hustling to build data centers. Compute is what everyone is hoarding. You guys must use a lot of compute. So talk about how you are using that AI and how that's limiting if you don't have the funds to make it happen. Yeah, we're using compute in a lot of ways. As you can imagine, when you're trying to generate scores for all 4 ,000 drugs versus all 18 ,000 diseases. I'll share the anecdote that when we built our very first version of this two and a half years ago, it took 100 days to generate a score for every drug versus every disease.

36:12The 75 million possibilities. Now it takes 17 hours. That's amazing. So there's been a lot of progress. That really is not because of work every year's done. That's because, as you know, just the incredible revolution that's occurring right now with artificial intelligence. But so it's a tremendous amount of compute. 17 hours for every time we run the 75 million. We run it multiple times a month to then come up with what we call a tier three run, which is the best one that the med team then sees the results of. We're able to assess based on known treatments, off-label treatments, just how accurate a particular set of predictions are.

36:42That's what the med team then reviews. And so there's a lot of iteration, which is so cool about having a 17-hour turnaround time so you can actually iterate a bunch of times. So we're both using compute to generate the 75 million scores. we're also using a lot of generative AI and additional LLMs to help us with triaging and coming up with packets around why something might be effective or not effective. And agents that ask questions like, is this the right use for this drug? What's the safety profile going to look like? And so all in, we're spending hundreds of thousands, probably low million a year on compute costs.

37:14So these are, and in many cases, we're actually saying to ourselves, let's scale back on this. Let's not actually run that extra analysis here. So that way we're not spending millions of dollars on compute. I don't want to be presumptuous, but Reed, you know something about this? Anything that can be done? Well, I mean, look, this is just stunning, I mean, in terms of the work you're doing. And it is exactly the kind of thing where people need to see the hope in AI and that it makes a difference in people's lives, literally and fundamentally. So, you know, before this podcast airs, we will move a million dollars to your compute budget.

37:53Oh my gosh. Ari, this is incredible. Because this, life-changing. Wow. Life-changing. Well, that is incredible. Thank you so much. That's amazing. No, I hope everyone knows about the work you're doing and like getting this accelerated. Because look, part of the thing, just to be clear, we've got thousands of drugs that we know are not toxic. Yes, exactly. That's the whole point of the 15-year thing. If they can cure another hundred lives, let alone thousands, tens of thousands, oh my God, literally it's what life changing means. It's what we're all here for. It's amazing. Thank you so much. That's incredible.

38:29My honor. That's incredible. Thank you so much for it. I mean, I think it also speaks to something else. Like so often when we're talking about AI, a lot of people want us to slow down. And I am so, I understand that point of view so much. Like the world is changing. Is your job going to be the same? What is happening? And for folks like you, you want AI to speed up Because if you could take that 100 days to 17 hours to one hour and have it be a tenth of the price or a hundredth of the price, imagine what more we can do. And so, again, just thinking about how AI can speed up what people like you are doing, that's what we're here for.

39:08So thank you both. And that's why we're doing this podcast too. Because it's literally to see what technology can do to make human life better. Exactly. What do you think the key thing to do here to help people understand how AI can really help medicine? Yeah, I think this is the right question. I mean, I think that if you use what we've done as a microcosm, at EveryCure, we are looking across all drugs, all diseases to find the lowest hanging fruit. Basically, the opportunities for humans to help people with the drugs we have. And we're looking at 75 million possibilities that our lab used to be able to look at dozens a year.

39:44and now we can look at 75 million. If you think about that and then just move that to other aspects of healthcare and you think about, okay, let's get away from this sort of very siloed, this doctor for this disease and you're going to have to wait months and months to find that person and they might not be the right doctor for you, you can imagine there could be this tremendous amount of efficiencies that are created for matching you with the right doctor at the right place to get the right drug and the right diagnosis. I've been so impressed by the research being done around, I think it's called ambient listening, were basically tools that can look at patients when they walk in.

40:17And maybe not just listening, but also visual. He does a lot of this. And they walk in the doctor's office and start coming up with diagnoses before the patients talk to the doctor. Just talking to the nurse, hearing the tone of their voice, seeing their gait, that's the kind of thing that I think that I'm so excited about in health care. I think that I personally, as a patient, struggled with getting diagnosis. I struggled with there weren't treatments for me. I think about sort of my whole journey and how it's so many levels. rapid diagnosis, find a drug that could be repurposed to treat that patient, make sure that that drug's working, follow biomarkers.

40:48I think that as you go through the healthcare system, I think there's just endless possibilities to improve our outcomes. This podcast is supported by Google. Hey, everyone, David here, one of the product leads for Google Gemini. If you dream it and describe it, VO3 and Gemini can help you bring it to life as a video, now with incredible sound effects, background noise, and even dialogue. Try it with a Google AI Pro plan or get the highest access with the Ultra plan. Sign up at Gemini.Google to get started and show us what you create.

41:25You guys are also promoting radical transparency. You've talked about releasing the data, the scores for millions and millions of pairs of drugs, and that could help other researchers, other folks who are doing the same thing. It could also help patients who are sort of looking for their cure. Talk about that transparency and maybe any pitfalls it could have and sort of what effect it could have on patients. Sure. So we only started every cure to help people. I mean, that's like we started this just because like we, you know, we just want to help people. There's drugs out there. There's people suffering.

41:55Let's help people. And we're really guided by that's our first core value is patient impact. And that's just like it's just our North star. It's why we're here. It's what we're doing. So when patient impact is your North star and your number one value, you have to optimize for, you know, what are the best ways to help patients. And so we felt from the beginning that one of the best ways to help patients is to make this information available rather than sort of keeping it within the organization, really, really to share it publicly. But it's important, we believe, to do it in the right way and do it responsibly.

42:30I that you talk about a lot. It's like, you know, let's push the frontiers of AI, but let's be thoughtful about how to do it responsibly. And so we're thinking a lot about how do we release them. We're planning to do it in early 2026. And so releasing with, you know, for example, disclaimers or training information, you know, like you have to go through a train and understand that like this is research data. You know, and actually even our team at EveryCure, we don't go from that drug scored the highest in our algorithm to let's give it to patients. We study in the lab. We look in clinical trials.

42:59is like there's a process to it. And so I think that sharing while also doing it in the right way, I think is going to be the best way to maximize patient impact. So there's so much hope and positivity in the transparency piece. And you sort of touched on the, and there could be misinterpretations. And I also feel like for, you know, for patients who want to have hope, there's some devastation. And oh, wait, they said this works for Castleman's, but it doesn't work for me. Or oh, I can't use it because they haven't. And so on the flip side, can you talk about challenges, misuses, misinterpretations of the data.

43:31You see people, you know, who have ailments, they're going on Reddit, they're diagnosing themselves, and then not necessarily good things happen. So what's the flip side? Yeah, you're absolutely right. I mean, this is an example of, I think, a perfect example of a double-edged sword, right, where, you know, on the one hand, information is critical, and doctors need it, researchers need it, patients need information. And so providing information is so important. And then the also important piece is that, to your point, there's a lot of information that's already out there. You know, when we provide scores, this isn't the first time that someone's going to see that a drug might work for their disease in a research setting.

44:05You know, the information's on PubMed all the time. But at the same time, the challenge would be that patients would maybe travel to another country to get a medicine that scores highly in our algorithm when that drug doesn't work because our algorithm makes predictions for drugs that are just not going to work. You know, the way that we think about it and the way that we've shown our platform works is that things that score towards the top are much better than things that score towards the bottom. But just because it's number one on this disease list doesn't mean it's better than number 10. Like, you know, like there's directionality.

44:35It's a candidate directionality. The higher it is, the better, but like 10 might actually be better than one. And, and you know what, for a lot, and this is a really important point for a lot of diseases, there actually isn't a repurposed drug that could treat that disease. I mean, I am so optimistic about all the patients we can help read, you know, we can help so many people, but we have to appreciate and understand that I don't personally believe that every human disease out there can be treated with every drug we have. What I do believe is that if there is a drug for that disease, we should find it, but not the vice versa, that there is a cure for every disease.

45:08And so the worry is that when people hear, to your point, Castleman's and Poem syndrome and angiosarcoma and data too, and these diseases where we've transformed the disease course with existing medicines, And it's like, well, then maybe it'll be for my disease. And I hope so bad. And we're trying so hard. And if it's there, we're going to do everything to find it. But there are certainly diseases where there is just not going to be an existing medicine that's going to treat that disease. We do need new drugs for those diseases. We absolutely do. Well, and also, I mean, part of that also why 10 and 1 is like we have a rough, as per Castleman's, a rough categorization.

45:43Yes. Well, which version of Castleman's do you have? Exactly. Maybe one is good for you and 10 is good for the other person. Exactly. And so you really have to understand this stuff. And it's one of the reasons why research and understanding is critical. So you've described giving these off-patent drugs. And we obviously, as you mentioned earlier, have a challenge that there's no market incentive for this. So how do you envision closing this gap? How do we get this to become like an amplified global phenomena? Yes. This is such a good question, and it's something we've been thinking about so much lately.

46:21We set up building this amazing tech team to make predictions. We built this amazing med team that identifies the best treatments and then does laboratory work and trials to prove they work. But to your point, just proving something works is not enough. We've got to get it to all the patients who can benefit. So we're actually now going to be building out a third team called our impact team, which is, okay, all the things that work, let's actually get them to all the people who can benefit. One part of it, as you mentioned, could be through an FDA label change. Another part of it is simply awareness raising.

46:47It's like talking to doctors, getting updates made to treatment guidelines. It's letting patients know and patient groups know. And so that third piece of this I think is so critical. And actually one of our board members, Bob McDonald, when we were talking about adding this third piece on, he said, well, David, yes, impact is the quote-unquote last leg. But impact needs to be infused throughout the first two. And by having an impact team, embedding them with our tech team and our med team, that will help us to create this circuit that we need. Is there something that should change on the systems level?

47:18Like should the U.S. government be the one funding this? Like should there be a different profit incentive? Is there something else maybe outside of your purview even that could change, whether it's in regulation or government or I don't know? A few things you mentioned I think are absolutely the kinds of things we should be thinking about. So one of them is government funding for our work. We're fortunate to be a recipient of funding from ARPA-H, you know, this brand new or few-year-old government agency. They've been amazing to work with over the last couple of years. And so I'm excited about continuing to work with ARPA-H.

47:52So that's one part. I think that should be part of it. I think, you know, I hope it will continue. Working closely with others outside of ARPA-H, you know, whether it's FDA, NIH, I think really thinking about this as a sort of a closed circuit where we're all working together, we're iterating, we're improving. I think that's a real opportunity for the future. And then to your point around what sort of policy changes could be made, the problem we're facing is that, as you mentioned, 80 % of drugs are generic, which means that there's no incentives to find new uses for them. But it's great that they're generic because that means that they're inexpensive and they're available around the world.

48:27But how do you create a system so that drug companies or anyone is incentivized to find new uses for these medicines when if they do find a new use, there's 10 manufacturers and they're selling it for a penny a pill. So I think it's going to require a lot of thought. I don't know if either of you guys have thoughts on this. I mean, how do you fix this challenge? Well, I mean, it's, I think, a public good challenge. So government, your very first answer, I think, is a good answer. The truth of the matter is that by the time that a drug gets off patent, the drug companies have made a ton of money off it, which they should.

48:58And that is the right way to recycle an innovation. and being a co-founder of a drug manufacturing company in its very early stages, obviously that economic incentive is very important. So you do the detailed research work, which really matters, right? And then how does the information of, oh, this drug that's already FDA approved has a good percentage off-label use for this? How does that enter into the therapeutic knowledge set? Yeah, great question. Well, in the current system, and this is another, I guess, area where the system could be broken or fixed. Right now, it's very much a paper gets published in an academic journal.

49:44And then hopefully, you know, doctors read that journal. And then maybe they talk about it at a conference. And then it sort of spreads slowly. I mean, there are two websites that doctors go to a lot for this information. So one of them is a website called UpToDate, where individual experts update their particular page for their disease. And then the other one is Wikipedia. Google, of course, is the third one. So doctors are going to other sources beyond, you know, just the published literature. But there isn't like a real – there isn't – you would imagine that there was this sort of one database where it's like this – you know, I have this disease for this patient.

50:18You know, what are my treatment options and, you know, what sort of research-grade treatment? But there isn't a really clear database for that. and I'm hopeful that we'll help to push something forward with that impact team. Yeah, no, that would be great. And it's probably actually interesting also, I mean, one of the things that people in the tech community have started talking about is, you previously had search engine optimization, although Google, now you're going to have agent optimization. And one of the things you certainly know is going to happen is anyone who gets confronted with one of these diseases where their doctor kind of says, there is really nothing for you.

50:48Every single one of those people is going to go back and go, okay, they go, hey, what about this? Yeah, exactly. Because you're wrapping that back into it. And even outside of what we're doing at EveryCure, people are already doing that in ChatGPT right now. I mean, it's already happening. Yes. Actually, I'd be curious about, I today say, if you've encountered a serious medical thing, including getting something from a doctor, you should go to a frontier model, ChatGPT, CLAW, Gemini, Copilot. and you should ask a second opinion, right? Because it's just worth getting the second opinion because if the second opinion contradicts the first opinion in any serious way, then go get a third.

51:31It isn't trust the AI per se at the end of the day, but to cross-check. Am I giving dumb advice? Am I giving smart advice? Am I giving questionable advice? Your advice is exactly what I say, but it's usually about a second human, like a second doctor. I say almost exactly the same thing. Okay, you've got your opinion. Go see another doctor. And if it's the same, then you feel pretty confident. But if it's different, then you've got to find a third doctor. Literally what you said. But the only change to replace is that maybe it's not a second doctor. Maybe it's actually now a lot. Maybe what I would say is that still consider having another human afterwards.

52:05So maybe it's like – If you have access to one. Yes. That's a great point. And by the way, the agent is also there right now. You walk out of Dr. One's office and you go, so – Yeah. Exactly. That's a great point. I like it a lot. Perhaps dissemination of the possibility of cures is much more difficult globally. So how do you think about the global context versus the U.S.? Well, I'll use a really specific example. So we right now have nine active drug repurposing programs, 11 that have passed through our Science Advisory Council. Among those nine, there's one that is a drug called DFMO for a rare neurological condition called Bachman-Bopp syndrome.

52:40So only about 20 kids have ever been described with this horrible neurological condition. and there's really really interesting data that was actually generated by dr bachman or dr bupp who the the disease was named after that a drug that was made for sleeping sickness called dfmo could potentially be effective for these kids who are born hypotonic they basically are on feeding tubes they're bedridden but if you give them df dmfo early on in their in their lives and the earlier you give it the better these kids can start to sit up and they can have a feeding tube taken out and they can walk. I mean, this is a drug for sleeping sickness that's really only used in one part of the world, but actually you could use it for these kids with Bach and Bup all over the world.

53:22And the way that it works is really, really fascinating. It really, it binds to the exact enzyme where there's too much in this particular condition. And so this is sort of a reverse of what you're describing. It's a drug that's actually widely available in many parts of the world, but now, you know, it could be useful in different places. And so I think that we're going to find more and more of these. And it's even, you know, when we got started, I think that my ingoing hypothesis for, well, where are you going to find drugs and diseases to repurpose? Ingoing, I would have said mostly autoimmunity and cancer, and in part because there's so much data on both of those categories of conditions and there's a lot of drugs that already exist.

53:56So I thought there would be a lot of matching there. I would have probably said that rare monogenic neurological conditions are probably going to be the hardest place to go. I mean genetic conditions where every single cell in the entire body has the same genetic mutation. These are enzyme deficiencies. These kids are bedridden. I would have said probably not there. Well, one of our first nine is one of these horrible, rare. And so like, I don't know where we're going to go. This discussion gives me an idea and I'm not sure if it exists or not. But it strikes me that either it should exist or it does exist, which is the FDA should have like, as it were, a new label process that should obviously be expedited.

54:38because you already know that you're dealing with something that has gone through the intense toxicity checks that exist. And, you know, there's some threshold of it's X percent effective, like 20 percent in Castleman's or something, but like, you know, some kind of thing and with some little detail on it. Does that exist and should it? So there is a there's there's two processes that exist. Neither of them are perfect fits for what you described. So one of them is called the 505b2 pathway, where basically if there's a drug that's approved for one thing and you want to change the dose or the formula of it, you can rely on all that prior data.

55:13So, like, you know, you already know that sildenafil, for example, was approved for erectile dysfunction. Researchers thought that it could work for a rare pediatric condition where kids weren't getting enough blood flow to their lungs. And so they said, but let's use a lower dose. So it's the same drug that could rely on all the previous work on sildenafil, lower dose. You could get approval for pulmonary arterial hypertension. So that's 505b2. if you change the dose or the formulation. If you want to use the exact same dose or the exact same formulation, there's theoretically a path forward for that.

55:40But realistically, if it's the same dose or the same formula, and you've created this new drug, let's say it's called like Casolimus, which is like serolimus for Castleman disease, but it's the same dose, same formula as before. And now your doctor writes a prescription for Casolimus. The pharmacist will say, well, serolimus is the same drug. And it costs less because it's a generic medicine. And so that will probably, the old drug will probably get utilized. So all that to say, you can go through these paths, but because the economics don't add up, no one goes through that step of changing the label for the same drug, for the same dose or formula, because no one's ever going to prescribe that same dose, same formula version.

56:21And up until now, FDA has interpreted the law as only the person who made the drug originally. The original innovator is allowed to submit to them, to ask for them to add a new disease to their – because the idea, at least in our current system, is that the drug company who made it the first time really owns that drug, even when it goes generic. Like, they're the quote-unquote innovator. But what we're working with FDA and what we're excited about is the potential, well, what if every cure could come in and we could change the label? Because we don't care if it's profitable. We're non-profit. Like, we want to go after non-profitable stuff.

56:53You know, we're setting a disease of 10 kids that have 20 kids that have been found, right? So we want to be able to go after them, And I'm hopeful that we'll come up with a path with FDA. That makes a lot of sense. So one of the things you've been talking about is you have this amazing technology. You find the drug that matches the disease. But then it doesn't matter if no one knows about it. So can you talk about you have a recent TED Talk. I think your book is being made into a movie. Talk about those promotional activities because, again, getting the word out is so critical. Yeah, well, you're absolutely right.

57:23And being on a podcast like this is amazing. Thank you seriously for this incredible opportunity. getting the word about what we're doing with every cure is just incredible um so yes so i i gave a ted talk and so we're excited for the reach that'll come from that um in in that ted talk we share that anyone listening can really help in any one of three ways one is that a lot of us have received drugs off label you know drugs that weren't intended for our disease and maybe it's helped or it hasn't helped let us know about it every card or slash ideas like you can actually contribute and we'll actually look to see how is that thing scoring or not scoring another is donating towards our work these trials are expensive the work's expensive the compute's expensive although you're You're going to help us out with the commute, which is amazing.

58:00So, you know, supporting us at everycard.org slash donate. And the third, to your point, is around raising awareness. Like getting the word out about DFMO to help us identify kids with Bachman-Bopp syndrome. Like a tweet, a social media post could actually help us find a parent of a kid with Bachman-Bopp who could get this medicine and they could, you know, take their feeding tube out, right? Like these are the kinds of things where that's not traditional, right? The traditional way is like get an FDA label change and we want to go through that process. But if we already have compelling evidence, the drug's available, it might be like a tweet or a social media post away.

58:34So I think it's helpful to know what some of these kind of repurposing stories. So share with us a couple of the ones that have been magical in your journey. Sure. So the first one that comes to mind was I told you about when it took 100 days for us to generate all of our scores. So this is two and a half years ago. So we ran it for the first time, and not surprisingly, maybe to you, one of the first diseases I looked at is Castleman's. Obviously, it's the disease I study. I wanted to see what scored highly. And the number one drug, the number three drug, were both TNF inhibitors, which had never been used before in Castleman's, but they're used for related immune-related disorders.

59:14And at the time, there was a patient who was very, very ill in Vancouver, and his doctor had been calling me every couple of days, you know, can we try this, can we try that? We tried basically all the drugs we'd ever used in other patients. We tried. They weren't working for this patient. And so I eventually recommended a TNF inhibitor to Luke, this patient's doctor, and said, really early days, like, you're telling me that you're about to transfer him to hospice? He said, yes, I've actually already written the order for a transfer to hospice. He's got a young daughter. He's got a wife. Like, they want to try something here.

59:45So he prescribed the TNF inhibitor, and amazingly, this patient's been in remission for over two and a half years now. We actually published an article on this in the New England Journal of Medicine earlier this year as this example of utilizing artificial intelligence to identify a drug that could help a patient. We also had some really interesting orthogonal laboratory data that helped us to feel more confident, because to your point, it's like you can't just trust AI alone. But when AI has an orthogonal data point that really points to it, that's when you can help people. So that was, for us, so special.

1:00:14And I was actually just emailing with that patient recently and hearing about how well he's doing. It's just it's everything. So that's one example that sticks with me. And then I'll give you two other quick ones because they're top of mind. One of them is this young girl named Kyla, who I mentioned briefly earlier, was dying in the ICU. She received this repurposed drug. And then just recently she was texting me. She just finished up summer working as this nursing assistant because she's in the middle of training to become a nurse, which is just so cool. Because of her experience, she wants to now take care of patients.

1:00:50And the third one, I mentioned the condition angiosarcoma. It's a horrible form of cancer, which is really a uniformly fatal form of cancer. And we utilized a drug called pembrolizumab for a patient named Michael for his angiosarcoma. at the time no one lived longer than a year with angiosarcoma but interestingly and this gets to sort of the power of artificial intelligence three years before we gave him this drug there was a paper published pointing to potentially a drug in that class maybe being helpful but it was a laboratory-based study had never been given to humans these are the kinds of things that sort of fall through the cracks when we can't keep track of billions of papers right like that you just can.

1:01:34But artificial intelligence can help us to make these connections. When I showed you those points in my knowledge graph, like, like those sorts of things substantiate knowledge graph edges. And so Michael got that drug pembrolizumab. Last year, he walked his son down the aisle eight years after he was supposed to die from his disease. Last weekend in Nashville, he walked his daughter down the aisle nine years after he was supposed to die from his disease. And so these are the kinds of things. And what you'll find as you're hearing me talk about these success stories is that I love it when we save someone's life, extend life.

1:02:08That's amazing. But what I really love, what we all love is to hear what they're doing in this extra time. We call it overtime, like time they didn't think they would have. They're walking their kids down the aisle. They're going to school. This for me is everything. All right. So rapid fire, the questions we ask all of our guests just to have a kind of a fun index. The first one is, is there a movie song or book that fills you with optimism for the future i'm a huge fan of adam grant you mentioned adam recently i think adam's book most recent book hidden potential provides me a lot of optimism and it's because it's all about the untapped potential that's already around us and of course you can imagine i told adam i was like i thought this was a book about drug repurposing when i when i saw the titles of hidden potential this is about drug repurposing it's not about drug repurposing but it is about the concept that there's so much around us that can help us that we actually, in many cases, already have the tools.

1:02:58We just have to find them. And that sort of concept that sometimes we describe as horizontal thinking, which isn't like let's dig deeply into finding a solution in front of us. Let's look horizontally, I think, to me, makes me hopeful that maybe there's a solution that's already out there for all the horrible things that we're facing in our world. Question number two, what is a question that you wish people asked you more often? You know what? It's a question that both of you guys have already asked me today, and it's how can we help. And I'm very fortunate. There are amazing people like both of you that say, how can we help?

1:03:30And it's the kind of thing like we actually need all of us to come together to help all of us. And I think we've touched on some of them, which you can either highlight or do a new. But where do you see progress or momentum outside of your field that inspires you? I mean, outside of our specific field of drug development, I mentioned that I'm really excited about some of this sort of ambience. work that can be done in healthcare that can help to diagnose patients. I mean, I just remember the struggles that I went through and I've seen for so many people to not get a diagnosis that I love the idea of putting artificial intelligence to work, to find diagnoses, to make diagnoses, even when we aren't sort of looking for them.

1:04:12Can you leave us with a final thought on what you think is possible to achieve in the next 15 years if everything breaks humanity's way, and what's the first step to get there? I think 15 years from now, we're being diagnosed with conditions before they're even showing any sort of signs, even based on data points that maybe we can't conceptually even pick up as humans. So diagnoses are coming in rapidly. Drugs are being, you know, delivered to your doorstep. Santa Claus theory of civilization is happening. You know, you're getting the drug right when you need it for the condition that you have.

1:04:44And that we, you know, massively reduce human suffering. And that, you know, I think about all those, you know, kids I sat with at AMF support group meetings. I think about the suffering I saw my mom go through and all the patients that we try to take care of. The idea of a world where we're able to relieve that suffering with the drugs that we have is just be amazing. So Reed, I'm looking at you to make sure that everything breaks the way we need to with artificial intelligence. We're trying. We're trying. And we can get to that world. Awesome. David, having you on the podcast makes me so, so happy.

1:05:15Thank you so much for being here today. Thanks so much for having me. This is awesome. Awesome. Possible is produced by Wonder Media Network. It's hosted by Aria Finger and me, Reid Hoffman. Our showrunner is Sean Young. Possible is produced by Katie Sanders, Edie Allard, Tanasi Dilos, Sarah Schleid, Vanessa Handy, Aaliyah Yates, Paloma Moreno-Gimenez, and Melia Hagudelo. Jenny Kaplan is our executive producer and editor. Special thanks to Surya Yalamanchili, Sayida Sepieva, Ian Alice, Greg Beato, Parth Patil, and Ben Relis.

1:06:14So try out Gemini 2.5 Flash at Aistudio.Google.com and let us know what you built.