The Decibel - Why AI is no miracle drug

Episode Date: September 17, 2024

Drug development has always been a long and arduous process, taking years of research and costing millions of dollars. When some biotech companies started to use artificial intelligence as part of that process, it was seen as a tool that had the potential to revolutionize drug discovery. Ten years on, those companies are faced with a reality check. Globe business reporters Joe Castaldo, who covers AI, and Sean Silcoff, who reports on  technology and life sciences, are on the show to talk about the promise of AI in drug development, and why the bets on technology haven’t panned out. Questions? Comments? Ideas? Email us at [e-mail]

Artificial intelligence is everywhere these days. We've seen innovations in self-driving cars, the advent of chat GPT and GPT-4, and advances in facial recognition technology. And some companies have even been betting on AI to revolutionize drug development. Getting a new medication to market typically takes years of research and millions of dollars. The thinking was, bringing in AI could drastically reduce that time frame. But a decade in, things aren't looking so promising. So today on the show, we've got Globe Business reporters Joe Costaldo, who covers AI, and Sean Silcoff, who covers technology and life sciences. They'll talk about the revolution that researchers and investors hoped AI would bring to the world of drug development and the reality that they're facing now.

I'm Mainika Ramanwelms, and this is The Decibel from The Globe and Mail. Joe, Sean, thank you so much for being on the podcast. Thanks for having us. Thanks. So Joe, you have been reporting on AI for years now. When drug developers were starting to kind of seriously use AI, what was the hope there? Like, what did they think AI could do? I think the hope of AI in drug development 10 years ago is really the same as it is today. Developing a drug is a long 10-year process that can cost a billion dollars and often ends in failure. So using AI was supposed to make this process faster, easier, cheaper, you know, result in new medicines, maybe that wouldn't have been discovered otherwise. And, you know, perhaps lower drug costs down the road to, yeah, just unlock insights that maybe human

scientists wouldn't have arrived at otherwise. Okay. Yeah. And on a very high level, like what kind of AI are we talking about here, right? Because there's all kinds of different things that AI can do. But what exactly is it doing here? Well, any AI model is trained on data, it looks at data, and tries to discern patterns in that data to make predictions. So with a chatbot, for example, you know, it's seen the whole internet and it can find patterns in language in order to be able to predict what the next word in a sentence might be. Which is what ChatGPT does then, of course.

Yeah. Correct. Yeah. And like have it in the right context. So the sentence makes sense. In drug development, it's the same idea. It's a model that looks at data and makes some kind of prediction. So as a quick example, Google DeepMind has a model called AlphaFold. And it's been trained on protein structures and sequences. And it can take a sequence of amino acids,

which make up proteins in our bodies that do all kinds of important functions. So it can take this sequence and predict what that protein will look like. What is the 3D structure of this protein in our bodies, which is not something we know easily. There's over 100,000 proteins in our bodies, which is not something we know easily. There's over 100,000 proteins in our bodies. And so that is actually very useful for drug developers to know that. Okay. And I guess in order to understand where this fits in, maybe we can just take a step back and kind of go over the basics of how drug development usually works. What are researchers

trying to figure out when they are developing a drug? Well, with genetic diseases, for example, there are a lot of things that a scientist would want to know. They'd want to know where is the genetic mutation, like in a string of DNA or RNA, where is the mutation? And then how does that mutation affect our bodies? What does it do in our bodies? How does it affect the proteins in our bodies that keep us alive and functioning? Is this gene being expressed too much or not enough? If you're going

to develop a drug, what protein are you going to target with your drug? Again, what does that protein look like? What is the structure of that protein so that you can help identify a drug molecule that's going to glom onto it to treat disease? Is it going to work? Is it going to be safe? Or is it going to have all kinds of side effects? So the questions really just go on and on. So it sounds like it's in those early stages then where AI could potentially help this

process. Yes. You know, in more traditional means, you know, you answer those questions by doing experiments in a lab or in a clinical setting, which takes time and costs money. The promise of AI there is to be able to make predictions in the early stages of things to help eliminate some of the guesswork and experimentation and dead ends that you might find otherwise. And those predictions still have to be validated in the real world, but it might get you closer to success a little bit faster. Sean, do you want to add anything there?

Yeah. I mean, when you think about sectors that have seen a lot of innovation, biotechnology has really taken to new approaches over the last 40 years or so. You've seen things like gene editing, creation of antibodies, cell therapies. And so artificial intelligence was fairly well accepted as a new and interesting pathway. You know, we're still learning a lot of things about the human body. Can't forget that biology remains a great mystery, even to the most sophisticated researchers in the world. So I think that's why AI was greeted with a lot of fanfare, a lot of hype, you saw a lot of a lot of investor money go into this. So can we look at specifically how this works? Like,

is there an example that we can talk through to understand how AI helps with this process? Sure. So we could look at a company called Deep Genomics, which is based here in Toronto, was founded in 2015 by Brendan Frey, who's a professor at the University of Toronto. He's very established in the Canadian AI world. He worked with Geoffrey Hinton, who's, professor at the University of Toronto. He's very established in the Canadian AI world. He worked with Geoffrey Hinton, who's one of the most famous AI researchers there is. And together, when Brendan was a PhD student under Hinton, they produced some work, important work that kind of lays the foundation for generative AI today. And so at some point in his career, he turned to medicine and biology,

and he wanted to use machine learning to answer this very complicated question of, if you have a string of DNA, like your own unique personal genome, can you use machine learning to identify and predict the consequences of a genetic mutation. So if there is a genetic mutation, can machine learning tell you what it's going to do in the body? Is it going to lead to disease? And if so, what and how? And that's a very complicated and ambitious question to answer. Sean, do you want to jump in, add anything here? So I followed the deep genomics story for years. And Canada Pension Plan, this is one of only two biotech companies in Canada that CPP put money behind. So this had a lot of fanfare behind it.

We followed their story. They identified some drug targets they wanted to go after. And I think what really kicked this story off for us earlier this year, I had an update call with Brendan and he kind of revealed that things had not gone the way they had hoped, not even close. AI has really let us all down in the last decade when it comes to drug discovery. Really, we've just seen sort of failure after failure. Do we have a sense of how much money has actually been poured into this space? I know it's easily into the billions of dollars. I mean, deep genomics alone has raised about 238 million US. And people we talked to said, you know, 10 years ago, if you said, hey, I'm going to use AI to solve drug discovery, you pretty much had money thrown at you. And certainly that was the case in the mid-2010s.

You saw that just about any company that had the right people and used that magic set of letters, A and I, could raise an astonishing amount of money. And, of course, we've seen that more recently, again, with the rise of generative AI. And we've seen a company in Canada, Cohere, raise hundreds of millions of dollars. And OpenAI is valued at tens of billions of dollars. So the space easily drew billions of dollars. Okay, so there's a lot of money at stake here. But let's go back to the deep genomics example that we were talking about a little bit earlier, because you mentioned that things didn't exactly go as planned. But Joe, what were they trying to do?

So the way that deep genomics went about this is they developed some 40 different AI models, and each one would work on a narrow slice of this drug development problem in terms of understanding genetics, understanding disease, developing molecules to treat the disease. And like, so they'd have, you know, an AI model to predict like the toxicity of a molecule, another one to predict, you know, how many patients you would need if you're running a clinical trial to show this drug is effective and the company was was very proud of us like brendan you know told us he used to say that other companies using ai are you know using ai for just one problem one drug development problem whereas um deep genomics is doing it all like

they're using ai for everything and And they did make progress. So they did identify a handful of molecules for different diseases that they wanted to take into clinical trials a few years ago, one of which was for a condition called Wilson disease. People who have it, their body accumulates copper, they can't process copper. So it builds up in their liver and the brain and other organs. And if it's not treated, it can be fatal. It's a rare condition. So, you know, the company was enthusiastic about this a few years ago, you know, they put out a press release saying that this is the first AI discovered therapeuticiscovered therapeutic candidate.

But when it came time to go into trials, they realized, well, there was a problem. One of their AI models made a mistake. It predicted a smaller number of patients than they would actually need to show that this drug was effective. And a similar thing happened with another one of their drug candidates. This AI model that predicts sort of patient size got it wrong. So this is like the number of people involved in a trial, that kind of thing? Correct. Yeah. And maybe that wouldn't have been fatal for the project, if not for the fact that Deep Genomics had a deeper problem than that. The larger problem is that having 40 different machine learning models is turned out

to be impractical. This is sort of the big mistake that Brendan talked to us about. So if you have 40 different models, each working on like a narrow slice of the same problem, they might not work well together. They might make predictions that contradict each other. It's a bit like that parable about, you know, blind men and the elephant. Like when one man, you know, touches the elephant's tail and he's like, oh, this is a rope. Another blind man touches the elephant's leg and thinks this is a tree trunk, right? Nobody is getting the full picture here and they're being led astray. The other issue is it's very hard to improve 40 different models. Like they have to manage, you know, 40 different data sets. There's different computing platforms. The way you improve AI models is through what they call scaling,

which is simply like pumping more data into it, throwing more computer chips at it, and it will improve. It will be more accurate. Deep genomics couldn't easily do that because again, they had to do that for 40 different models. So it was completely impractical in the end. It's like herding 40 different cats. Instead of one cat. Yeah, instead of one big cat. We'll be right back. So we've kind of talked about the broad issues that they're facing then, but maybe we can get into some specific challenges here then in using AI for drug development.

So what can't AI do that I guess researchers were hoping that it could do? So there are a few big challenges that I think, you know, all companies are facing when it comes to AI and drug development. One, as we've referenced before, is just biology is a very hard, gnarly problem. Somebody told us it's not like physics where there are well-understood rules and laws. There are still a lot of unknowns. And it's challenging to build an AI model to solve a problem that you yourself don't understand. Data is another big issue. So, you know, as we said before, every AI model needs data to learn from, you know, today's generative AI models have been trained on, you know, all of the internet, essentially. But for drug development, data doesn't necessarily exist in sufficient quantities

or, you know, in good enough quality to solve some of these problems. Like the cliche with AI is it's garbage in, garbage out. If you have bad data, you're going to get bad results. So developing good quality data to train these models, it takes time and costs money and has held some companies back. Yeah. And what about the people involved in this kind of work? Like what is the kind of expertise, I guess, needed, really, in such a complex field? Yeah, I think there's something that comes from years and years of training in university and lab settings, that teaches you how to be a biologist teaches you how to be a chemist teaches you how to be a drug researcher. And you can't just be an expert in one field like machine learning and then plunk into

biotech and and understand it from scratch i mean it takes years of understanding that process and dealing with all these machines and and the various bits and pieces that you research there was one of the people we spoke to who said you, you can't just hire a bunch of machine learning people who work to skip the dishes, or some other consumer internet company, they might have been mighty good at machine learning and artificial intelligence. But you can't just move those people into the biotechnology laboratory setting, and expect them to replicate their expertise to the same effect. Yeah. So we've been talking a lot about this company, Deep Genomics. There must be other companies doing similar research. Like have they, I guess, encountered similar failures?

How have they dealt with that? For sure. Yeah. Some companies that use AI in a big way have brought drug candidates into clinical human trials in recent years. And in some cases, it hasn't gone all that well. So there's a company called Benevolent AI out of the UK that was running trials for a drug they had developed for eczema. And last year, when they announced the results of these trials, it turned out that the drug didn't really reduce itch and inflammation, which is what you would want out of an eczema drug. There's another company called Sumitomo out of Japan that had a drug for schizophrenia that was developed in part with the use of AI, also in human trials, and it didn't do better than a placebo they found. And so, you know,

in the case of Benevolent, like after these results, they actually cut the eczema drug program, they laid off some workers, they reduced their lab footprint. They announced a merger with another kind of AI drug developer. And, you know, both their stock prices have been falling for many years. In the case of Sumitomo with the schizophrenia drug, they said earlier this year that it's going to be challenging to generate any revenue from this. And they ended up licensing out development to another company. So effectively, you know, washing their hands of it. And to be clear, like these kind of failures in trials is not unusual in drug development, like drugs fail all the time. But it just shows that, you know, AI is not the magic bullet.

And, you know, there are some problems that it just cannot solve for yet. And I think that's one of the things that really makes AI in drug discovery so interesting is you have these very binary outcomes with a hard and fast adjudication process, which is the drug regulators, like the Food and Drug Administration from the US, it must pass their test to move forward. And the FDA kind of brings down the hammer. And that's that. And I think that probably, for the most part, sets AI and drug discovery apart from many of the other potential applications in other fields. Can we can we talk about that? Like, what is the part of the process then that they managed to use AI for where it actually is helpful? Yeah. So there's a company called Ventus Pharmaceuticals that has offices in Montreal and Boston. And so we talked about how the lack of data is a real challenge. So this company has

kind of figured out a workaround. So they're looking at like protein binding, like how do you get a molecule to bind to the protein you want to target? And that's a huge challenge because there are a lot of proteins. They have different shapes. They have multiple binding sites. Those binding sites can have different shapes, sizes, chemical properties. So there's just like a mind-bogglingly high number of permutations here. And we just don't have enough data on protein binding sites for AI to be of much use, according to Ventus. So they're taking a different approach. Proteins are inside of cells in our bodies. There is water inside of our cells, and that water

sloshes around. It bumps up against these protein binding sites. And so Ventus uses AI to model the behavior of water inside these protein binding sites that then gives the company like a blueprint essentially of what these binding sites look like, which helps the company narrow down molecules that could bind to these proteins. Like it doesn't have to test, you know, thousands of molecules. It could narrow it down to dozens. That doesn't mean that that dozen molecules, they're going to bind, but they have a greater likelihood. So it can save time and money. The CEO kind of characterized it. He had this poetic way of putting it. If you want to understand the universe, you can try to model and map every object there is in

the universe, which is infinite. Or you can study the behavior of light in the universe to give you a bigger picture of what's going on. Yeah, that's really interesting. It sounded poetic when he said it. It sounded poetic when you said it too. It actually really does let us kind of understand how the concept is working. So that's really interesting. Just in our last few minutes here, I mean, we've talked about the challenges of AI and drug development. I wonder what lessons can maybe other fields learn about how AI was applied to drug development and what worked and maybe what didn't and how they could do something better. In the short term, I think a lot of new technologies get overhyped and that feeds a hype cycle.

And then the hype cycle quite often crashes. And then in the long term, the technology ends up delivering sometimes all of the promise that it was built on. And I think what we very clearly discovered is that AI has utility in biology and chemistry and drug discovery. Nobody's saying it doesn't, and nobody's saying it won't work. What we are saying and seeing clearly is it hasn't worked nearly as well as people thought it would be to justify the immense amount of investor money that poured into the space, the big promises that were made by some of these companies, and the output that we've seen so far. There has yet to be a molecule designed by AI like that magic potion that has come to

market. That's not to say that won't happen, but the odds are particularly high and are partially based on the shortcomings of the data that we don't know, the things we don't know about biology. So I think, you know, it's a good reminder that tech and innovation move us forward and real progress is made, sometimes not at the cadence in the time and based on the big promises people make. But that doesn't mean that you should necessarily dismiss or discard the whole thing. You do so at your peril. I think if anything, innovation is full of perils.

Believing the hype is a peril and dismissing the hype is also a peril. Joe, anything to add there? You know, direct parallels or lessons to other sectors is a bit tricky because drug development is unique. But, you know, it's a bit of a reality check for AI. I mean, new AI applications can seem amazing at first and can seem like magic, but then you scratch the surface and they're more brittle than they appear at first. The same thing has happened with self-driving vehicles. There was a lot of hype about autonomous

cars, billions of dollars invested, but it's proven way more difficult and way more expensive to get self-driving vehicles than people anticipated. So, you know, it's a reminder that, you know, AI isn't a magical cure. It can still be very powerful, but it's worth keeping expectations in check. Joe, Sean, thank you both so much for being here today. Thank you. Thank you. That's it for today. I'm Mainika Raman-Wilms. Today's episode was edited and mixed by Allie Graham. Our producers are Madeline White, Michal Stein, and Allie Graham. David Crosby edits the show. Adrian Chung is our senior producer, and Matt Frainer is our managing editor.

Thanks so much for listening, and I'll talk to you tomorrow.