How AI Is Changing The Drug Development Pipeline
Researching and developing new drugs is a notoriously long and expensive process, filled with a lot of trial and error. Before a new drug gets approved scientists must come up with something they think might work in the lab, test it in animals, and then if it passes those hurdles, clinical trials in humans.
In an effort to smooth out some of the bumps along the road, a growing number of pharma companies are turning to new artificial intelligence tools in the hopes of making the process cheaper and faster. Ira talks with Will Douglas Heaven, senior editor for AI at MIT Technology Review about his reporting on the topic.
Will Douglas Heaven is the senior editor for AI at MIT Technology Review. He’s based in London, England.
IRA FLATOW: This is Science Friday. I’m Ira Flatow.
Researching and developing new drugs is a notoriously long and expensive process. It’s also a process filled with a lot of trial and error. Before a new drug gets approved, scientists must come up with something they think might work in the lab. Then they test it in animals. And then, if it’s passed those hurdles, they test it in humans.
In an effort to smooth out some of the bumps along the road, a growing number of pharma companies are turning to new artificial intelligence tools in the hopes of making the process more efficient. Yes, AI is here also.
Joining me now to talk about his reporting on the topic is my guest, Will Douglas Heaven, senior editor for AI at the MIT Technology Review. He’s based in London. Welcome to Science Friday.
WILL DOUGLAS HEAVEN: Hey. How are you doing?
IRA FLATOW: Nice to have you. Let’s talk about that. But before we get into AI and how it’s changing drug development, can you give us a quick refresher on how new drugs are typically designed and developed?
WILL DOUGLAS HEAVEN: Yeah. So I mean, this would be super high level. You need to identify something in the body that your drug is going to interact with. And that’s called a target. And then you need to obviously design the drug that’s going to interact with that target. And, I don’t know– say your target is a protein in the body.
Then you want to design a drug molecule that will find its way to that target and either change how it functions or switch it off. And this is also typically done in the very early stages in the lab. You put it in a sort of chemical soup and see that it does what you hope it does. And the promising ones then go down– as you mentioned, further down the pipeline and get tested in animals and eventually in humans.
But all of that is really slow and painstaking, with loads of dead ends.
IRA FLATOW: We’re talking about using AI to speed up this process, right, before it gets to human testing?
WILL DOUGLAS HEAVEN: Exactly. So the big promise with AI is that you can cut some corners, do less of the work in the actual lab– which is expensive, takes a lot of people a lot of expertise, and experiments are slow to run. And so the idea generally is, if you can do as much of this in a computer, with AI-driven simulations, predicting which molecules will do what with which targets– if you can do all that in a computer and avoid the dead ends before you actually then do the expensive lab work, then hopefully you can do the whole thing faster and cheaper.
IRA FLATOW: OK. Let’s start with something practical we can talk about. How about identifying a new target for a new medication. How are researchers using AI to improve this process?
WILL DOUGLAS HEAVEN: I mean, it’s all drawing on the vast amounts of data that we’ve acquired over the last few years. So sophisticated computational techniques are by no means new. And one of the reasons that AI is starting to make a big impact now is that that data, there’s enough of it and it’s high enough quality that you can use it to train AI algorithms that can make predictions about what this molecule might do with that target in this sort of biological situation.
And you can run many, many, many of these simulations all at once, searching through vast spaces of potential interactions– orders of magnitude more than anything we’ve ever been able to do before. And hopefully, you can then highlight the one in a million-billion that is promising. And then, only make that and test that actually in the lab.
IRA FLATOW: Let’s talk about how drugs work in the body. You have a drug. It’s a molecule. And it has to fit into a certain place in a certain spot on the cells, is that correct?
WILL DOUGLAS HEAVEN: Yeah.
IRA FLATOW: And so it’s like a Tinker Toy. It’s really a shape thing, right– a lock and a key thing sort of thing? And so you’re looking for the right shape of the right molecule to fit it in the right spot. And that seems to be really tedious and maybe something AI can do a lot better.
WILL DOUGLAS HEAVEN: So if you have lots and lots of data about what kind of keys are out there, what kind of locks are out there, and what they might do when they go together, then you throw all that at your machine learning system and it will learn the patterns of locks and keys. And when you ask it a question about a particular lock, it can predict whether or not this particular molecule will be a good fit for it.
IRA FLATOW: Can you give an example of how pharma companies are using AI image generation?
WILL DOUGLAS HEAVEN: I mean, one thing that you might want to do is actually take some tissue sample from patients. So that might be tiny bits of cells. And then you train your robotic camera on them, and apply lots of these different potential drug candidates to those samples and see what happens.
And the computer vision can monitor in fine detail, all right, there’s some changes that happened– tell whether that drug is killing the wrong cells or not doing anything at all. So it just allows you to do experiments at scale. So you don’t need actual human eyeballs on these experiments all the time. You can run them automatically in a big robotic lab. Again, it just speeds things up, and does it with far more accuracy than maybe an individual human could.
IRA FLATOW: You also featured a company called Exscientia, which uses AI to match patients with drugs that they might otherwise not have been recommended. This is terrific.
WILL DOUGLAS HEAVEN: Yeah, that’s very cool. I mean, everything we’ve been talking about so far is from the bottom up, when you’re talking about how you might identify new targets and then build new drugs that would act on those. They’re also looking at things from the other end, from the patient end, where you might want to actually figure out exactly which drug matches that patient.
So again, you would take a tissue sample from that patient. And then, in the lab, with these robotic computer vision, you can then test lots of different existing drugs– try 100 of them– on this patient’s cells and see which one actually works.
I mean, you would never try 100 drugs on a patient. Think of someone who has to go through chemotherapy. That’s an extremely unpleasant, drawn out process. And if it doesn’t work, all you’re able to say is, OK, that drug we just tried wasn’t any good for that patient. So let’s try the next one on the list. And it might take you months or years to go down half a different drugs. And if they’re all not working, then that’s an awful thing to put a patient through.
IRA FLATOW: Right. And you profiled in your work and your reporting a patient that has actually benefited from this technique.
WILL DOUGLAS HEAVEN: They’ve run a trial now on quite a few different patients. But the first one that really gave them a sense that they were on to something was an 82-year-old patient, who had been through six failed chemotherapy sessions. So six different drugs that his doctors had tried on him hadn’t worked. And that’s months and months of suffering through all this.
So his doctors didn’t really have anywhere else to turn and nothing else to lose so they enrolled him in this new trial. They took tissue samples from the patient. They tested dozens all at the same time. And the amazing thing there is that the drug that they found that actually worked that they then gave to the patient was a drug that was on the market already, but its previous tests had suggested that it wouldn’t be any good for the patient they were looking at.
IRA FLATOW: No kidding.
WILL DOUGLAS HEAVEN: That’s why the patient’s doctors hadn’t actually tried it on him. But it goes to show that every person is different in complicated, subtle different ways. So just because the tests had shown that this drug hadn’t worked for most people or on average, it doesn’t mean that it wouldn’t work for that particular person. And in this case, it did. But they wouldn’t have found it if they hadn’t done this AI-driven robotic lab.
IRA FLATOW: That’s amazing. Because that’s the definition of personalized medicine right there.
WILL DOUGLAS HEAVEN: Exactly.
IRA FLATOW: And also what you’re doing is you’re probably– which is something drug companies are really concerned with– and that’s money– you’re probably making it cheaper to be able to do this, right, using AI?
WILL DOUGLAS HEAVEN: Yeah. And who benefits from that in the end, right? It’s us. The reason drugs are so expensive is because the cost of every successful drug has to cover the 19 or so drugs that didn’t make it. Estimates slightly vary, but 1 in 20 drugs might actually make it all the way through from initial development all the way through years of clinical trials to actually make it onto the market.
One person I spoke to in the industry said that, basically, the business of drug discovery is about failure. I mean, you expect new drugs to fail. If you can make the process quicker by avoiding dead ends to begin with, throwing away all the candidates that AI predicts won’t actually go anywhere, then the whole process becomes cheaper. And if you also only then submit to clinical trials those drugs which seem to be most promising, then you’ll have a higher success rate when it gets to clinical trial.
If we can make it so that the drugs that we actually make and develop and then put to clinical trials are more likely to be successful, then the cost of drugs for everyone should go down.
IRA FLATOW: Well, has this gone mainstream? Is it a big trend? Or are these just small startups that are doing the AI work and then big pharma will come in when it’s developed?
WILL DOUGLAS HEAVEN: Well, a bit of both. Most of the innovation and activity is happening in the startup space. Which is typical in tech, right? And there’s a lot of hype and a lot of money being thrown at these startups. I mean, one investor I spoke to reckoned there were several hundred pharma startups that have emerged in the last few years, all looking at different aspects of this, all wanting a bit of the action.
I mean, I expect there’s going be a correction in the next few years, where many of those won’t make it. But the activity in the space is really hot. And obviously, a startup can move quicker and try things and fail faster. But the big existing pharma companies, they’re not blind to this. I mean, they’re also trying out these techniques themselves.
IRA FLATOW: Well, when do we know when all of this is a success? You follow this quite well. Is there a timeline? Do you see a progression when, well, maybe next year, a year from now, this will be mainstream and everybody will be using it?
WILL DOUGLAS HEAVEN: It will probably be a few years before we see the first drugs that were designed with the help of AI actually hit the market.
IRA FLATOW: The mark of success of drug development is when you start clinical trials, right? I mean, when will we see AI drugs be part of clinical trials?
WILL DOUGLAS HEAVEN: There are a bunch of drugs already that have been submitted to clinical trials. They’re in that stage now where they’re being tested in humans– around about 20. I mean, there will be more admitted every few months. But let’s say around about 20 drugs are now in clinical trials, where they’re being tested on humans. And that’s up from zero in 2020. So in the last couple of years, that’s just one measure of how fast this is progressing.
Now, clinical trials can run for years. So it might be some time before we see the first successful ones actually hit the market, actually be allowed to– for regulators to actually allow your doctor to prescribe it to you. And, of course, we may find that the drugs that have been submitted in the initial round, maybe none of them work. So we might go back to scratch, and then these companies will have to try again and submit new drugs.
But even if that happens then, this technology is not going away. I mean, the upsides in terms of speeding things up and cutting costs, I think, are too great.
IRA FLATOW: Yeah. And AI is here everywhere to stay. Why not in the drug industry, Will? Thank you for taking time to be with us today.
WILL DOUGLAS HEAVEN: No, it’s my pleasure.
IRA FLATOW: Will Douglas Heaven, senior editor for AI at the MIT Technology Review. He is based in London.