#3 - Beautiful Mistakes: The Serendipity of Drug Repurposing

WEBVTT
00:00:00.761 --> 00:00:04.671
So asking for a friend how was Viagra discovered?
00:00:05.740 --> 00:00:07.267
It was a beautiful mistake.
00:00:17.021 --> 00:00:18.085
Hello and welcome.
00:00:18.085 --> 00:00:21.407
My name is Vasant Sarathy and I'm Laura Hagopian.
00:00:21.407 --> 00:00:23.225
We are very excited today.
00:00:23.225 --> 00:00:28.550
Hi, laura, hi, we are going to be talking about drug repurposing.
00:00:29.960 --> 00:00:40.375
Yeah, I'm excited for this topic because Viagra is just one example, but it's about finding new uses for existing drugs.
00:00:41.679 --> 00:00:42.343
Very cool.
00:00:42.343 --> 00:00:43.787
Yeah, so let's get started.
00:00:48.840 --> 00:00:49.180
Where do we begin?
00:00:49.180 --> 00:00:50.143
Well, first of all, why even do this right?
00:00:50.143 --> 00:00:52.470
Why not just like discover new drugs?
00:00:52.470 --> 00:00:56.421
The main reasons are that it saves time and money, right?
00:00:56.421 --> 00:01:09.207
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.
00:01:10.209 --> 00:01:35.284
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.
00:01:35.284 --> 00:01:39.212
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.
00:01:39.212 --> 00:01:51.593
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.
00:01:51.593 --> 00:01:56.930
It's much easier to use a drug that's already in supply, if that makes sense.
00:01:57.831 --> 00:02:00.927
Yeah, that does, and it seems like a really smart thing to do.
00:02:00.927 --> 00:02:04.631
Now the obvious question is will this drug even work, right?
00:02:04.631 --> 00:02:12.211
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.
00:02:13.379 --> 00:02:13.479
Right.
00:02:13.479 --> 00:02:21.006
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.
00:02:21.006 --> 00:02:23.651
But let me tell you about what happened with Viagra.
00:02:23.651 --> 00:02:35.128
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.
00:02:35.128 --> 00:02:44.307
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.
00:02:44.307 --> 00:02:56.051
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.
00:02:56.051 --> 00:02:58.044
Can you guess what that side effect is?
00:02:59.669 --> 00:03:00.651
Extended erections.
00:03:02.040 --> 00:03:06.181
Yeah, it treated erectile dysfunction and that's not what they were testing the drug for.
00:03:06.181 --> 00:03:14.401
So it was sort of a beautiful mistake, right, and it's almost like it was a shot in the dark.
00:03:14.401 --> 00:03:16.165
They weren't even trying to find it right.
00:03:16.165 --> 00:03:22.002
It was just sort of serendipitous that it happened and that's great and all.
00:03:22.002 --> 00:03:34.471
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?
00:03:34.471 --> 00:03:46.457
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.
00:03:46.478 --> 00:03:58.427
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.
00:03:58.427 --> 00:04:10.110
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.
00:04:10.919 --> 00:04:11.060
Right.
00:04:11.060 --> 00:04:22.994
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?
00:04:22.994 --> 00:04:26.982
Right, and you could look at things like genes, like biology, like 3D structure.
00:04:26.982 --> 00:04:30.911
But I'm talking all theory here, and this is what scientists do.
00:04:30.911 --> 00:04:35.228
Right, this is what they do to try to find drugs that they can repurpose.
00:04:35.228 --> 00:04:46.392
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?
00:04:47.175 --> 00:04:50.122
Yes, and so that's actually yes, so I'm really.
00:04:50.122 --> 00:05:05.853
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.
00:05:05.853 --> 00:05:27.730
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.
00:05:27.730 --> 00:05:51.387
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.
00:05:51.939 --> 00:06:00.071
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.
00:06:00.071 --> 00:06:06.982
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.
00:06:06.982 --> 00:06:19.271
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.
00:06:19.271 --> 00:06:20.793
And so let's you know.
00:06:20.793 --> 00:06:21.694
With that in mind, let's, let's.
00:06:21.694 --> 00:06:24.276
Are you interested in learning a little bit about knowledge graphs?
00:06:24.415 --> 00:06:50.569
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.
00:06:51.052 --> 00:06:54.658
That's what's going through my head right now, yeah, and so let's, let's take that apart now.
00:06:54.658 --> 00:07:01.641
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.
00:07:01.641 --> 00:07:04.649
So, at the most fundamental level, a knowledge graph.
00:07:04.649 --> 00:07:16.345
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.
00:07:16.345 --> 00:07:30.339
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.
00:07:30.339 --> 00:07:35.310
And, uh, the arrows represent relationships between those things.
00:07:35.310 --> 00:07:41.872
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.
00:07:41.899 --> 00:07:45.230
I was like, oh gosh, I have this picture in my head, but yes, please, please, please.
00:07:45.230 --> 00:07:45.750
An example.
00:07:45.980 --> 00:07:48.266
Let's take the example of movie recommendations.
00:07:48.266 --> 00:08:06.108
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.
00:08:06.108 --> 00:08:08.305
So how would this be represented in a knowledge graph?
00:08:08.305 --> 00:08:17.333
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.
00:08:17.333 --> 00:08:27.363
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.
00:08:27.363 --> 00:08:29.528
Now, already we have three types of nodes.
00:08:29.528 --> 00:08:31.940
We have an, a user node, for alex.
00:08:31.940 --> 00:08:34.609
We have a genre genre and we have a movie and the.
00:08:35.532 --> 00:08:41.235
Once you have that relation, once you have those nodes set up, you can now draw arrows between them to establish relationships.
00:08:41.235 --> 00:08:50.465
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.
00:08:50.465 --> 00:08:56.270
So it basically represents the fact that Alex likes sci-fi, right, likes the sci-fi genre, okay.
00:08:56.270 --> 00:09:05.311
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.
00:09:05.311 --> 00:09:09.402
So Inception, the movie, has the genre of sci-fi.
00:09:09.402 --> 00:09:14.452
So now you have three circles connected by two arrows that represent several facts.
00:09:14.452 --> 00:09:27.892
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.
00:09:27.892 --> 00:09:29.885
Right, you already have this knowledge.
00:09:29.885 --> 00:09:38.451
Now the question is, in a movie recommendation situation, you might be interested to know if Alex would like some other movie.
00:09:38.879 --> 00:09:40.580
Yeah, that's exactly what I was thinking.
00:09:41.481 --> 00:09:46.511
And so this graph in theory doesn't have an edge maybe from, let's say, interstellar.
00:09:46.511 --> 00:09:53.715
That's another movie, that's a Christopher Nolan movie, that's also sci-fi, but Alex has never watched it.
00:09:53.715 --> 00:09:55.961
And you're wondering would Alex like this movie?
00:09:55.961 --> 00:10:02.172
And there's obviously a circle or a note for Interstellar, right, because it's another thing that you're representing.
00:10:02.172 --> 00:10:17.119
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.
00:10:17.330 --> 00:10:19.535
Because it has the genre sci-fi Right.
00:10:19.797 --> 00:10:22.841
Yes, right, and it has the genre sci-fi.
00:10:22.841 --> 00:10:37.178
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.
00:10:37.198 --> 00:10:51.700
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.
00:10:51.700 --> 00:10:58.336
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.
00:11:00.870 --> 00:12:09.144
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.
00:12:09.909 --> 00:12:13.273
And then, can you explain to me how link prediction is that right?
00:12:13.273 --> 00:12:17.999
Number one and number two can you explain to me how link prediction would fit into that what I just said?
00:12:18.340 --> 00:12:19.572
Absolutely so.
00:12:19.572 --> 00:12:20.537
That is exactly right.
00:12:20.537 --> 00:12:36.804
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.
00:12:36.804 --> 00:12:43.823
You can ask the question does drug A or could drug A treat disease B?
00:12:43.823 --> 00:12:56.946
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.
00:12:56.946 --> 00:13:00.360
So there's a good candidate there, potentially.
00:13:00.360 --> 00:13:07.342
And finding those is for a human eye, it's like finding a needle in the haystack it's really hard.
00:13:07.591 --> 00:13:08.676
Well, that's what happened with Viagra, right?
00:13:08.730 --> 00:13:11.554
It was just sort of like yeah, you got really lucky Right and this like well you got really lucky.
00:13:11.575 --> 00:13:14.821
yeah, right, and this like well, here I guess you're making your own luck.
00:13:15.763 --> 00:13:25.274
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.
00:13:25.274 --> 00:13:27.176
You're no, look, you have all this data.
00:13:27.176 --> 00:13:36.946
Why don't we use AI and techniques in computer science to mine links in this knowledge graph?
00:13:36.946 --> 00:14:05.125
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.
00:14:06.070 --> 00:14:08.816
Yeah, okay, so like a human basically could not do this.
00:14:08.816 --> 00:14:12.724
It's just like too big of a problem to wrangle.
00:14:13.029 --> 00:14:20.256
Yeah, and CKG is one of them and it's, you know, has 16 million nodes and it in fact has 220 million edges.
00:14:20.256 --> 00:14:25.001
So you're predicting the 221st millionth edge.
00:14:25.001 --> 00:14:29.287
In that sense, because you're predicting a new edge given what's out there already.
00:14:29.287 --> 00:14:32.395
Sense, because you're predicting a new edge given what's what's out there already.
00:14:32.395 --> 00:14:33.099
And these things are very complex.
00:14:33.099 --> 00:14:35.572
It's not just diseased gene target, but some of these have many more types of nodes.
00:14:35.572 --> 00:14:43.677
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.
00:14:43.677 --> 00:14:48.614
So they get very complex, very, very fast and um, which is a good thing.
00:14:48.614 --> 00:14:58.734
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.
00:14:58.734 --> 00:15:09.217
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.
00:15:10.158 --> 00:15:15.376
Which, like you know, if you think back to Viagra, that was something that was serendipitous.
00:15:15.376 --> 00:15:17.416
It was a mistake, it was whatever you want to call it.
00:15:17.416 --> 00:15:22.719
We just happened to like walk into that drug and it turned out to be a very big success.
00:15:22.719 --> 00:15:29.337
But here you're talking more about having a systematic way to discover new uses for old drugs.
00:15:29.859 --> 00:15:30.600
Yeah, exactly.
00:15:30.600 --> 00:15:33.254
And so let's talk about that a little bit more, like how, exactly?
00:15:33.254 --> 00:15:37.916
If you have a knowledge graph, how would you then go about finding that link right?
00:15:37.916 --> 00:15:44.143
So there's a whole host of different techniques and there are pros and cons to each of these techniques.
00:15:44.143 --> 00:16:08.282
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?
00:16:08.330 --> 00:16:09.796
There's like connections between things.
00:16:09.796 --> 00:16:18.639
How do you capture that in numbers to ensure that things that are connected are also somehow numerically connected, right?
00:16:18.639 --> 00:16:30.437
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.
00:16:30.437 --> 00:16:35.336
You want to make sure that neighborhoods in the graph, node neighborhoods, are maintained.
00:16:35.336 --> 00:16:40.076
You want to make sure that certain paths exist still and can be easily found.
00:16:40.076 --> 00:16:44.336
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.
00:16:44.336 --> 00:16:54.976
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.
00:16:54.976 --> 00:16:58.172
Okay, that is kind of black boxy if you.
00:16:58.211 --> 00:16:58.974
Yeah, that's what I was.
00:16:58.974 --> 00:17:12.161
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?
00:17:12.161 --> 00:17:19.653
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.
00:17:19.653 --> 00:17:55.174
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.
00:17:55.829 --> 00:17:58.398
Yeah, and so not only is that a problem.
00:17:58.398 --> 00:18:03.634
I think the other problem is, if you have an embedding of a graph, you're kind of stuck with that.
00:18:03.634 --> 00:18:06.259
It's really hard to update the embedding with a new node.
00:18:06.259 --> 00:18:16.920
You have a new piece of information, you have a graph that's constantly growing and you're learning new information every day.
00:18:16.920 --> 00:18:18.967
You want to have a graph that is flexible, and so you want to reason over a graph in that flexible way.
00:18:18.967 --> 00:18:20.772
So does that mean you have to retrain the whole system?
00:18:20.772 --> 00:18:24.711
Now, it might be the case you do, but there are techniques that people are getting around that.
00:18:24.711 --> 00:18:29.539
But you're right, the fundamental issue is still one of the black boxy nature of the system.
00:18:29.539 --> 00:18:33.693
So explainability is really important, and knowledge graphs have that already, right.
00:18:33.693 --> 00:18:34.276
So why not?
00:18:34.276 --> 00:18:35.137
Why aren't we using that?
00:18:35.137 --> 00:18:36.891
And I think so.
00:18:36.911 --> 00:18:54.125
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.
00:18:54.125 --> 00:19:01.155
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.
00:19:01.155 --> 00:19:09.252
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.
00:19:09.252 --> 00:19:25.090
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.
00:19:25.732 --> 00:19:30.066
there could be another one which is I likes maybe alex in this example that we had before.
00:19:30.066 --> 00:19:40.426
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.
00:19:40.426 --> 00:19:43.821
So finding that connection is easier now because we have this general rule.
00:19:43.821 --> 00:19:51.241
And having that rule is really useful because we can later say actually that's not true or we can explain it.
00:19:51.241 --> 00:19:57.103
We can sort of have an understanding of why we recommended this new other movie to Alex.
00:19:57.411 --> 00:19:59.357
Okay, but question who makes the rule?
00:19:59.357 --> 00:20:01.037
Is like AI making the rule.
00:20:01.037 --> 00:20:03.257
Am I looking at the data and making the rule?
00:20:03.257 --> 00:20:03.919
How does that work?
00:20:04.138 --> 00:20:04.681
That's a good question.
00:20:04.681 --> 00:20:10.383
You know the techniques that people use and it's sort of broadly called rule mining is it's done automatically.
00:20:10.383 --> 00:20:18.344
That is, they look at the overall patterns in the graph and try to figure out what rules are basically hold right.
00:20:18.344 --> 00:20:21.355
So you sort of come up with these if-then rules, kind of.
00:20:21.355 --> 00:20:25.813
You know, and there's a whole body of different types of rules that exist.
00:20:25.813 --> 00:20:38.336
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.
00:20:38.336 --> 00:20:44.337
But it's automatically generated and it's based purely on the existing patterns that already exist in the graph.
00:20:45.569 --> 00:21:01.801
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.
00:21:01.869 --> 00:21:12.097
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.
00:21:12.097 --> 00:21:13.101
So it's dynamic.
00:21:13.101 --> 00:21:27.391
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.
00:21:27.530 --> 00:21:28.594
Okay, yeah, yeah so.
00:21:28.854 --> 00:21:39.383
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.
00:21:39.383 --> 00:21:47.655
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.
00:21:47.655 --> 00:21:54.878
But the advantage of rules is that you then are able to have explainability significantly more explainability for a new recommendation.
00:22:02.329 --> 00:22:05.856
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.
00:22:05.856 --> 00:22:11.336
I want to be able to, like have an idea of why it would work.
00:22:12.560 --> 00:22:13.423
Yeah, I mean that's it.
00:22:13.423 --> 00:22:15.068
These are sometimes people call them meta paths.
00:22:15.068 --> 00:22:15.509
Yeah, I mean that's it.
00:22:15.509 --> 00:22:18.588
These are sometimes people call them metapaths, that are a path through the graph.
00:22:18.588 --> 00:22:24.196
Normally, a path through the graph is just kind of walking from one node to another along the edges, along the arrows.
00:22:24.196 --> 00:22:26.829
Metapath in this case is more general, right?
00:22:26.829 --> 00:22:40.637
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?
00:22:40.678 --> 00:22:43.616
I mean that's at the end of the day right, you still have to do the testing, right, yeah?
00:22:43.616 --> 00:22:51.593
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.
00:22:51.593 --> 00:22:57.070
It's just like now you have an idea of what drug could work and it's not finding a needle in a haystack.
00:22:57.070 --> 00:23:08.050
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.
00:23:08.050 --> 00:23:34.435
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.
00:23:37.140 --> 00:23:40.163
Awesome, awesome, this I'm telling you.
00:23:40.203 --> 00:23:41.726
This is one of my favorite topics.
00:23:41.726 --> 00:24:16.717
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.
00:24:16.717 --> 00:24:19.477
They've read every single thing on the internet.
00:24:19.477 --> 00:24:52.183
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.
00:24:53.090 --> 00:24:54.073
It's explainable, right.
00:24:54.073 --> 00:24:54.734
Well, it's explainable, right, like.
00:24:54.734 --> 00:25:04.243
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.
00:25:04.243 --> 00:25:14.054
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?
00:25:14.054 --> 00:25:15.777
That's key.
00:25:16.679 --> 00:25:18.201
Yeah, yeah, absolutely.
00:25:18.201 --> 00:25:20.192
There's a ton of research on this.
00:25:20.192 --> 00:25:25.615
We will share a lot of that in the show notes, but I want to leave us with sort of a few takeaways here.
00:25:25.615 --> 00:25:44.438
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.
00:25:44.438 --> 00:25:46.759
But now we can do it systematically and intentionally.
00:25:46.759 --> 00:25:54.730
So we can go from having sort of lucky anecdotes to systematic data-driven drug discoveries.
00:25:55.030 --> 00:25:56.936
With the help of AI, even more right.
00:25:56.936 --> 00:26:03.922
Right right, because you could do knowledge graphs without AI, but with AI it's like even bigger, exactly.
00:26:05.951 --> 00:26:16.000
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.
00:26:16.000 --> 00:26:19.801
It's not just a matter of the AI telling us an explanation.
00:26:19.801 --> 00:26:25.623
I mean, you can ask Chad GPT for an explanation, but that is not true to how it arrived at its answer.
00:26:25.623 --> 00:26:33.859
And in this particular instance, the explainability is really solid for knowledge graphs because you know the rule that was underlying its decision making.
00:26:34.300 --> 00:26:47.603
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.
00:26:48.683 --> 00:26:56.199
Yes, and I think the third takeaway, what I love, love, love about this particular application is the fact that it's so multidisciplinary.
00:26:56.199 --> 00:27:14.192
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.
00:27:14.192 --> 00:27:25.222
Going forward with AI research and with, you know, applying technology to human use cases is making sure all the stakeholders' voices are heard.
00:27:25.222 --> 00:27:26.545
I think that's really key.
00:27:27.049 --> 00:27:32.259
Good thing we have, like you know, an AI expert and a doctor on a podcast together to talk about these things.
00:27:32.259 --> 00:27:32.800
That's right.
00:27:33.823 --> 00:27:34.223
That's right.
00:27:34.223 --> 00:27:36.936
Yeah, I think that's all we have for you for now.
00:27:36.936 --> 00:27:41.698
If you found this episode enlightening, please go ahead and hit subscribe.
00:27:41.698 --> 00:27:48.400
Tell a friend, tell a colleague who cares about it, and we'd be happy to hear your questions.
00:27:48.400 --> 00:27:50.555
We'll see you next time, thank you.