Code & Cure
Decoding health in the age of AI
Hosted by an AI researcher and a medical doctor, this podcast unpacks how artificial intelligence and emerging technologies are transforming how we understand, measure, and care for our bodies and minds.
Each episode unpacks a real-world topic to ask not just what’s new, but what’s true—and what’s at stake as healthcare becomes increasingly data-driven.
If you're curious about how health tech really works—and what it means for your body, your choices, and your future—this podcast is for you.
We’re here to explore ideas—not to diagnose or treat. This podcast doesn’t provide medical advice.
Code & Cure
#3 - Beautiful Mistakes: The Serendipity of Drug Repurposing
What if the next breakthrough treatment for a rare disease was already sitting on the pharmacy shelf?
Drug repurposing, the science of finding new uses for existing medications, is transforming how we discover treatments, blending serendipity with strategy. It began with surprises like Viagra, a heart drug turned blockbuster, but today it's driven by advanced data tools that accelerate discovery and reduce risk.
We explore how knowledge graphs (vast maps of biomedical relationships between drugs, genes, and diseases) are now at the core of this revolution. When paired with artificial intelligence, these networks can surface overlooked connections buried in decades of medical literature. Unlike opaque algorithms, these AI systems can explain why a drug might work for a new condition, providing testable hypotheses and building trust with clinicians.
This approach doesn’t just save time—it can save lives. Traditional drug development takes over a decade and billions of dollars. Repurposed drugs, having already passed safety checks, can reach patients faster and cheaper. That’s a game-changer for rare and neglected diseases where time and resources are limited.
This episode is a journey through beautiful mistakes and brilliant methods, showing how multidisciplinary teams, from data scientists to clinicians, are reshaping the future of medicine. Join us to learn how technology is turning chance into choice, and uncovering new hope in old drugs.
References
Drug repurposing: approaches, methods and considerations | Elsevier
Elsevier Industry Overview
(No individual author listed)
Trends and Applications in Computationally Driven Drug Repurposing
Luca Pinzi & Giulio Rastelli
International Journal of Molecular Sciences, 2023
Biomedical Knowledge Graph Refinement with Embedding and Logic Rules
Sendong Zhao, Bing Qin, Ting Liu, Fei Wang
arXiv preprint, 2020
COVID-19 Literature Knowledge Graph Construction and Drug Repurposing Report Generation
Qingyun Wang et al.
NAACL Demonstrations, 2021
Explainable Drug Repurposing via Path Based Knowledge Graph Completion
Ana Jiménez, María José Merino, Juan Parras, Santiago Zazo
Scientific Reports, 2024
Knowledge Graphs for Drug Repurposing: A Review of Databases and Methods
Pablo Perdomo-Quinteiro & Alberto Belmonte-Hernández
Briefings in Bioinformatics, 2024
Credits:
Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/
So asking for a friend how was Viagra discovered?
Speaker 2:It was a beautiful mistake.
Speaker 1:Hello and welcome. My name is Vasant Sarathy and I'm Laura Hagopian. We are very excited today. Hi, laura, hi, we are going to be talking about drug repurposing.
Speaker 2:Yeah, I'm excited for this topic because Viagra is just one example, but it's about finding new uses for existing drugs.
Speaker 1:Very cool. Yeah, so let's get started.
Speaker 2:Where do we begin? Well, first of all, why even do this right? Why not just like discover new drugs? The main reasons are that it saves time and money, right? So if you think about how much it costs to bring a new drug to market, it costs a lot in terms of time like 10 or more years sometimes and it can cost in the hundreds of millions or even billions of dollars to do it.
Speaker 2:And so what happens if you repurpose a drug is that you're reusing a drug that was used for something else, and so you already know that that drug is safe and that that drug is effective if it's been approved by the FDA already, and so you're able to more efficiently bring it to the market. It lowers the time and it lowers the risk of failure, because a lot of drugs drop out that are found to be not safe and not effective. And in addition to that, drug repurposing is a great option for rare diseases, because it's hard to put so much time and energy and investment into a rare disease. It's much easier to use a drug that's already in supply, if that makes sense.
Speaker 1:Yeah, that does, and it seems like a really smart thing to do. Now the obvious question is will this drug even work, right? I mean, there's one thing about side effects and getting FDA approval and making sure it's safe and all of that, but obviously we wanted to actually treat the disease.
Speaker 2:Right. So then the question becomes like what's the most effective way to repurpose drugs, and Viagra is actually not the best use case for that. But let me tell you about what happened with Viagra. So Viagra was developed by Pfizer in the 1980s and they were looking for new drugs to treat heart disease and they were thinking about what's the mechanism of action here. They were looking for drugs that would kind of dilate the blood vessels and help people with angina and other heart conditions, and so they put it into trials. It was in the clinical trials, where they were checking okay, is this safe, is this effective, et cetera that they found that a lot of the men who were using the drug reported a side effect. Can you guess what that side effect is?
Speaker 1:Extended erections.
Speaker 2:Yeah, it treated erectile dysfunction and that's not what they were testing the drug for. So it was sort of a beautiful mistake, right, and it's almost like it was a shot in the dark. They weren't even trying to find it right. It was just sort of serendipitous that it happened and that's great and all. But then the question becomes well, hey, are there other drugs out there that could be repurposed, and should we be looking for it in a more effective way? This was a mistake, but maybe if we look under the surface and we figure out ways to do this better, we could repurpose lots of drugs out ways to do this better.
Speaker 1:We could repurpose lots of drugs, yeah, and it seems like also thinking this way allows us to think about getting drugs on the market quickly, especially when there's an emergent situation, like we had in the pandemic, where we needed a drug out there fast. And if there exists a drug already that has the potential to work, then that it seems like the logical thing to do is to put that in the pipeline and get that moving, if it in fact works.
Speaker 2:Right. So you could look at things like, okay, what's the mechanism of this drug, how does it work, how does the disease itself work and what might we want to target to improve the condition? Right, and you could look at things like genes, like biology, like 3D structure. But I'm talking all theory here, and this is what scientists do. Right, this is what they do to try to find drugs that they can repurpose. But this is where I want to bring you into the conversation, because it is a difficult problem and it's interesting to think through, like could AI help here?
Speaker 1:Yes, and so that's actually yes, so I'm really. This is one of the reasons I'm really excited about this topic is because not only is it about AI, but it's about a kind of AI that I personally, I conduct research in and I'm also very excited about, which is this notion of symbolic representations. That is, the idea that you have AI systems that are often black boxes, but it would be cool if we were able to see inside them and make sense of what was happening, what the computer scientists say representations, what we mean is just pictures and English, and being able to understand what's being said and what's being reasoned over would be really helpful for us, and I'm very excited about this for that reason. And in the drug repurposing world, one of the key representations is something that's called a knowledge graph, and this is a form of writing down the various relationships between the drugs and the genes and the disease, so that we can just sort of capture all the knowledge we have about them and then think about them collectively.
Speaker 1:So one of the things that people do and I'll get into what Knowledge Graph is in a second but one of the advantages is that now you have lots of research papers out there. You have lots of clinical studies out there for the drugs and for various other sort of up-and-coming compounds that could serve as drugs as well. All of that is available publicly, and there's been a large effort in the AI space to turn those into knowledge graphs which we then can mine for new drugs. And so let's you know. With that in mind, let's, let's. Are you interested in learning a little bit about knowledge graphs?
Speaker 2:Yeah, and I, I totally am, and I think you're bringing this up and I'm like, oh gosh, if I were to create this manually, like that would be a lot of work and I feel like I would miss some of the paper, some of the literature, some of the databases out there, right, like that's where you know, could I construct this on my own, like, maybe, question mark, but like it wouldn't be as good as something that could synthesize from so many different locations all at the same time.
Speaker 1:That's what's going through my head right now, yeah, and so let's, let's take that apart now. Let's talk about knowledge graphs more, um, at a very fundamental level, and then we can work our way up to what would happen at scale. So, at the most fundamental level, a knowledge graph. Imagine, when I say knowledge graph, imagine a picture with a bunch of circles and arrows pointing from one circle to another, and now you imagine, and that's what's collectively called a graph, it's a mathematical term. Those circles are sometimes also called nodes and those arrows are sometimes called edges, and the circles represent things in the world that you want to have, you have knowledge about, and the or things in the world that you want to represent. And, uh, the arrows represent relationships between those things. And I think it's easy, uh, to conceptualize this with a more practical example okay, please, yeah, because I was just gonna ask that.
Speaker 2:I was like, oh gosh, I have this picture in my head, but yes, please, please, please. An example.
Speaker 1:Let's take the example of movie recommendations. We all like movies and let's say we have somebody called Alex who we know likes sci-fi movies and has watched Inception, the Christopher Nolan sci-fi movie, and we know this information. So how would this be represented in a knowledge graph? The way this would be represented in a knowledge graph is by having a node or a circle that represents Alex, because that's one of the entities we care about. There's a node or circle for sci-fi, because that's a genre we care about, and then a node or a circle for Inception, because that's a movie we care about, and then a node or a circle for inception because it's a movie we care about. Now, already we have three types of nodes. We have an, a user node, for alex. We have a genre genre and we have a movie and the.
Speaker 1:Once you have that relation, once you have those nodes set up, you can now draw arrows between them to establish relationships. So an example arrow would be an arrow from the circle for Alex to the circle for sci-fi and that, would you know, the arrow might have a name, it might be called likes. So it basically represents the fact that Alex likes sci-fi, right, likes the sci-fi genre, okay. And then maybe there's another arrow that goes from inception to sci-fi and that might be an arrow that says has a genre. So Inception, the movie, has the genre of sci-fi. So now you have three circles connected by two arrows that represent several facts. Now you can imagine this extending to thousands of users, maybe thousands of movies, maybe you know dozens of genres and you have arrows connecting all of them about the knowledge that you already have. Right, you already have this knowledge. Now the question is, in a movie recommendation situation, you might be interested to know if Alex would like some other movie.
Speaker 2:Yeah, that's exactly what I was thinking.
Speaker 1:And so this graph in theory doesn't have an edge maybe from, let's say, interstellar. That's another movie, that's a Christopher Nolan movie, that's also sci-fi, but Alex has never watched it. And you're wondering would Alex like this movie? And there's obviously a circle or a note for Interstellar, right, because it's another thing that you're representing. But because you have all the other information, you might be able to figure out a pattern about Alex's likes and dislikes and then determine and conclude that there's a likelihood, a good likelihood, that Alex will probably like Interstellar as well.
Speaker 2:Because it has the genre sci-fi Right.
Speaker 1:Yes, right, and it has the genre sci-fi. Now, it's possible, because of the complexity of the collection, it might be the case that although Alex likes Christopher Nolan's sci-fi movies, he might really dislike leonardo dicaprio right and so that's why he so.
Speaker 1:So he might, you know, uh, excuse me, he might dislike matthew mcconaughey, who's an interstellar, and so because of that, that's another fact that we have to play and take into account as well, but that might play out in other movies or other things that Alex has done as well. So you can imagine how complex this can get pretty quickly, yeah, from the sheer number of nodes and edges that we have in this graph.
Speaker 2:Yeah, so, with that, so I think it's worthwhile thinking about graphs, because computer scientists have been thinking about graphs for a really long time, and what I just described here is a task that's called link prediction or knowledge graph completion, where you're predicting a brand new uh, link or aspect of the knowledge graph that you care about okay, so I'm gonna play this back to you, but I'm gonna play it back in like science, health speak, yeah, so in if we were talking about drug repurposing, which we are, the nodes, the circles would be things like a drug or a disease or a gene or a protein or something like that, and then the edges or the arrows would be things like oh, treats this or interacts in this way or is associated with, and all of that data could be gathered from all sorts of different locations, right, like experimental results, maybe health records, the literature, databases, et cetera, in order to create this like giant representation.
Speaker 2:And then, can you explain to me how link prediction is that right? Number one and number two can you explain to me how link prediction would fit into that what I just said?
Speaker 1:Absolutely so. That is exactly right. And you have, you know, nodes for diseases, nodes for genes, nodes for drugs and so on, and because you have the relationships of things like treats or targets or is caused by the kinds of relations you might have between those nodes, you can start to predict new things. You can ask the question does drug A or could drug A treat disease B? And drug A might be known to treat disease A, but it might be the case that disease B is also caused by the same gene that drug A targets. So there's a good candidate there, potentially. And finding those is for a human eye, it's like finding a needle in the haystack it's really hard.
Speaker 2:Well, that's what happened with Viagra, right?
Speaker 1:It was just sort of like yeah, you got really lucky Right and this like well you got really lucky.
Speaker 2:yeah, right, and this like well, here I guess you're making your own luck.
Speaker 1:Yeah, I mean that's right, and so I think that's the big value about this notion of drug repurposing is you take the guesswork out. You're no, look, you have all this data. Why don't we use AI and techniques in computer science to mine links in this knowledge graph? Now, just to give you a sense of the scale of these knowledge graphs, there's been a lot of work actually that's been done in drug repurposing with knowledge graphs, building these knowledge graphs from data, so from papers and from clinical studies and such, and there's over a dozen different data sets and they range in sizes, just to give you a sense of scale, from, you know, 20, 30,000 nodes for this to some that have over 16 million nodes.
Speaker 2:Yeah, okay, so like a human basically could not do this. It's just like too big of a problem to wrangle.
Speaker 1:Yeah, and CKG is one of them and it's, you know, has 16 million nodes and it in fact has 220 million edges. So you're predicting the 221st millionth edge. In that sense, because you're predicting a new edge given what's out there already. Sense, because you're predicting a new edge given what's what's out there already. And these things are very complex. It's not just diseased gene target, but some of these have many more types of nodes. Uh, some of them have up to 60, 70 different types of nodes, covering everything from proteins to all kinds of things, um, in that space. So they get very complex, very, very fast and um, which is a good thing. I I'm arguing that it's actually a really good thing because now we have a representation, um, albeit hard to read, of all of this knowledge in this common, shared structure. So you might have papers across the board that have different ways of writing the same thing, but that's all been mined out, and now we have representations that are, you know, in a shared, unified structure which we can then then look at.
Speaker 2:Which, like you know, if you think back to Viagra, that was something that was serendipitous. It was a mistake, it was whatever you want to call it. We just happened to like walk into that drug and it turned out to be a very big success. But here you're talking more about having a systematic way to discover new uses for old drugs.
Speaker 1:Yeah, exactly. And so let's talk about that a little bit more, like how, exactly? If you have a knowledge graph, how would you then go about finding that link right? So there's a whole host of different techniques and there are pros and cons to each of these techniques. One set of techniques involves deep neural networks Our friend, our AI friend that we've been talking about a ton and these techniques involve taking the entire graph and finding a way to represent them as numbers while still retaining the structure of the graph, because remember that the graph itself has a visual structure to it, right?
Speaker 1:There's like connections between things. How do you capture that in numbers to ensure that things that are connected are also somehow numerically connected, right? And so there's a whole area of research and that's called graph embedding, where they try to take the graph and they put it into a numeric form, while still maintaining the mathematical properties of a graph that you have. You want to make sure that neighborhoods in the graph, node neighborhoods, are maintained. You want to make sure that certain paths exist still and can be easily found. You want to make sure so there's all of these structural aspects of the graph that then need to be encoded in the numbers. But once you have it in the number form, you can apply convolutional neural networks or other techniques to be able to ask questions about the graph, to do the link prediction tasks that you have. Okay, that is kind of black boxy if you.
Speaker 2:Yeah, that's what I was. I was getting a little nervous there because one of the things I mean, one of the concerns I always have as a clinician, thinking about how we use ai, is, like, you know, is this something that I can understand? Or, you know, there's like an input that goes in and an output that comes out and then, like what happened in between, and I like want to know that. I want to be able to explain that Sometimes I realize things are not explainable, but I really do appreciate if the AI, like reasoning, if you call it that is something that, like I could look under the hood of and be like oh yeah, like this drug might be repurposed for this reason because of this gene or this protein or the structure that's three-dimensional, like I want to be able to see that information, to be like oh yeah, this makes sense or this doesn't make sense, but if there's a black box there, then I don't get to see it.
Speaker 1:Yeah, and so not only is that a problem. I think the other problem is, if you have an embedding of a graph, you're kind of stuck with that. It's really hard to update the embedding with a new node. You have a new piece of information, you have a graph that's constantly growing and you're learning new information every day. You want to have a graph that is flexible, and so you want to reason over a graph in that flexible way. So does that mean you have to retrain the whole system? Now, it might be the case you do, but there are techniques that people are getting around that. But you're right, the fundamental issue is still one of the black boxy nature of the system. So explainability is really important, and knowledge graphs have that already, right. So why not? Why aren't we using that? And I think so.
Speaker 1:There's a whole host of other techniques that involve what's called rule mining, where what you're doing is, instead of just having it be end to end which is another term that people use all the time as in you given the graph and you answer with a link that might be a potential drug, that's end to end. The alternative would be rule mining, where you have an intermediate step where you extract some rule, some pattern from the graph in an understandable way. So, going back to our example of the movies, you can think of a rule that is sort of a let's call it a co-actor rule. That is, if I like, uh, leonardo dicaprio movies, right, that that's a rule, that that we have now, so if I like one, I might like another okay so you're able to apply this more general rule that I like these kinds of movies in the graph to find new connections.
Speaker 1:there could be another one which is I likes maybe alex in this example that we had before. Um, we've learned from looking at all the movies that he's watched that he actually likes sci-fi movies, so recommending another sci-fi movie to him would be a really good idea based on that. So finding that connection is easier now because we have this general rule. And having that rule is really useful because we can later say actually that's not true or we can explain it. We can sort of have an understanding of why we recommended this new other movie to Alex.
Speaker 2:Okay, but question who makes the rule? Is like AI making the rule. Am I looking at the data and making the rule? How does that work?
Speaker 1:That's a good question. You know the techniques that people use and it's sort of broadly called rule mining is it's done automatically. That is, they look at the overall patterns in the graph and try to figure out what rules are basically hold right. So you sort of come up with these if-then rules, kind of. You know, and there's a whole body of different types of rules that exist. But you basically come up, you have the AI system propose rules for itself and then run through the graph and see how likely that rule is to hold and how likely that rule is to not hold and then, based on that, you know, find different ways of ranking the rules and such. But it's automatically generated and it's based purely on the existing patterns that already exist in the graph.
Speaker 2:And so if we added a new person like Laura, who also likes sci-fi movies, does that, then you can make me follow those same rules, like oh, you think I'm going to like Inception and Interstellar because Alex does and because I like sci-fi, or something like that that's right.
Speaker 1:But remember that if that fails, as in you watch the movie and you hate it and somehow you're able to report back to the knowledge graph that you hate it, then that's added back to the graph and the rules are updated. So it's dynamic. Yes, so if the rules are updated, so it's dynamic, yes, so if the rules are in fact, uh, if that rule, specific rule that we just took into, uh, took into account about laura, it are people or the rule could be even more general people who like sci-fi movies, also like leonardo dicaprio movies, I don't know something like that.
Speaker 2:Okay, yeah, yeah so.
Speaker 1:But if that rule is very sort of there's lots of cases where that's true and lots of cases where that's not true, then a new piece of evidence that suggests it's not true is going to tip it more in that direction. Whereas rules that are more solid are going to be hard to turn around, and so finding exceptions are a little bit harder and there's a whole area of research around this. But the advantage of rules is that you then are able to have explainability significantly more explainability for a new recommendation.
Speaker 2:Yeah, see, I'm a fan of that, because I do want to understand why, for example, a new drug or an old drug could get repurposed for some new indication. I want to be able to, like have an idea of why it would work.
Speaker 1:Yeah, I mean that's it. These are sometimes people call them meta paths. Yeah, I mean that's it. These are sometimes people call them metapaths, that are a path through the graph. Normally, a path through the graph is just kind of walking from one node to another along the edges, along the arrows. Metapath in this case is more general, right? Because you're not just walking along one edge, you're saying making some generalization across the entire graph, and so having that rule allows you to have it be more transparent, more human, readable, and essentially it functions as a hypothesis, right?
Speaker 2:I mean that's at the end of the day right, you still have to do the testing, right, yeah? So I mean, yeah, you, you, you maybe could skip phase one trials because we already know it's safe and effective, but you're, you still have to test the drug. It's just like now you have an idea of what drug could work and it's not finding a needle in a haystack. It's like, oh, here's a systematic way we've developed to figure out how else this drug could work or for what other indication or what else it could treat, and now we need to go do trials. But you've already jumped through this big hoop that new drugs have to jump through, because we already know, if it's being repurposed, that it's safe and effective, and so you could shave time off by already having that, and you could shave basically money off the drug discovery process too, because you're able to skip a whole piece of the process yes, that's, that's right.
Speaker 1:Awesome, awesome, this I'm telling you.
Speaker 1:This is one of my favorite topics. I think the knowledge graph field is one that went through a lot of ups and downs during the early 2000s and 2010s, and people used it for all kinds of different purposes, but with the advent of LLMs and ChatGPT and all of these other tools, a lot of knowledge was encoded in these systems, and so then you started wondering what the value of knowledge graphs was, because these knowledge graphs were sort of constructed over the years and people thought that potentially, llms internal knowledge subsumes whatever is in those knowledge graphs, because, in theory, the LLMs are reading all the same papers. They've read every single thing on the internet. But, as we just talked about, I believe that there's a distinct value in knowledge graphs in that there is a sense of, there's a certain sense of what's the word I'm looking for Guarantee Assurance maybe is a better word that these knowledge graphs that are constructed from data were done carefully and have truth to them, and grounding our predictions on them is better than grounding our predictions on what an LLM might say.
Speaker 2:It's explainable, right. Well, it's explainable, right, like. I think you hit on that point earlier, but I do think this concept, especially when it comes to medicine and healthcare, is really important. It's not just like going from A to B, but like what brought you there and why would this drug work in this new situation? That's key.
Speaker 1:Yeah, yeah, absolutely. There's a ton of research on this. We will share a lot of that in the show notes, but I want to leave us with sort of a few takeaways here. I think we've covered quite a bit I mean, there's so much more here to cover right but I think there's sort of a couple of key takeaways, the first one being that repurposing is not the same as guesswork, and in fact we showed that with the example of Viagra, where we had the serendipitous encounter. But now we can do it systematically and intentionally. So we can go from having sort of lucky anecdotes to systematic data-driven drug discoveries.
Speaker 2:With the help of AI, even more right. Right right, because you could do knowledge graphs without AI, but with AI it's like even bigger, exactly.
Speaker 1:The second takeaway I think for us is really explainability is important, and we come back to it over and over again and it really matters. It's not just a matter of the AI telling us an explanation. I mean, you can ask Chad GPT for an explanation, but that is not true to how it arrived at its answer. And in this particular instance, the explainability is really solid for knowledge graphs because you know the rule that was underlying its decision making.
Speaker 2:Right, and that is what builds the clinical trust in something like this to say, okay, let's move forward, let's move this prediction into a clinical trial and see if this old drug can work for a new purpose.
Speaker 1:Yes, and I think the third takeaway, what I love, love, love about this particular application is the fact that it's so multidisciplinary. You have knowledge graphs, you have AI techniques, but those graphs are made up of specialized, expert knowledge about biology and medicine and they tie so closely with both the mathematics but also practical uses, and I think that this sort of captures a direction that we all should be doing. Going forward with AI research and with, you know, applying technology to human use cases is making sure all the stakeholders' voices are heard. I think that's really key.
Speaker 2:Good thing we have, like you know, an AI expert and a doctor on a podcast together to talk about these things. That's right.
Speaker 1:That's right. Yeah, I think that's all we have for you for now. If you found this episode enlightening, please go ahead and hit subscribe. Tell a friend, tell a colleague who cares about it, and we'd be happy to hear your questions. We'll see you next time, thank you.
