AI Was Supposed To Discover New Drugs. Where Are They?

IRA FLATOW: Hi. This is Ira Flatow, and you’re listening to Science Friday.

[MUSIC PLAYING]

Today on the show, experts have been promising new effective drugs designed by AI. So where are they?

PETER COVENEY: We have nothing actually on the market that can be attributed in any way directly to AI. So something has gone wrong. It’s overblown expectations.

[MUSIC PLAYING]

IRA FLATOW: With AI basically everywhere these days, there’s debate about how useful generative AI actually is. But one area where experts think it could be game changing is scientific research and speeding up major discoveries.

One of the fields that experts think is ripe for this application is drug discovery, an expensive and time-consuming process. It can take decades before a drug comes to market. But so far, there’s great disappointment.

The few AI-designed drugs that have made it to clinical trials haven’t been approved. And what was once a hot space for venture capital investment has cratered in the last few years. And many startups have shut their doors. So why exactly has it been so hard to make AI-designed drugs. What are the fundamental issues? And what does the future of this space look like?

Here with some answers is Dr. Peter Coveney. He’s the director of the Center for Computational Science in University College of London’s Chemistry Department. He studies how chemistry discoveries can be sped up with algorithms and computers. Peter, welcome to Science Friday.

PETER COVENEY: Thank you very much indeed.

IRA FLATOW: You’re welcome. Before we get into the challenges, can you walk us through normally how drugs are made the old-fashioned way? Why does it take so long? Why is it so expensive?

PETER COVENEY: It’s a big challenge. We would like to understand how the body works to design drugs correctly. We don’t have all that knowledge with us today. So we have to do what’s the next best thing.

And that’s to start with the most efficient way of getting a drug out that works for the maximum number of people. That’s the magic bullet idea that one drug fits all. And that’s the approach that the pharmaceutical companies adhere to today. Let’s try and get a drug that will apply to everyone.

That’s not likely to work very well because we are all different. So there’s this challenge between understanding how individuals react to drugs and the population as a whole.

So you will try to find the one molecule that you think fits into a target protein and then kills off some disease. That itself still takes a lot of time because you have to first be clear what the target protein is, and then the search for the magic molecule starts to take place.

Now you focus so much on that single property that it’s rather later on in the process that people start to worry about other aspects of how the drug candidate will work. There may be side effects to that drug. And then you put it through several stages of clinical trials.

And indeed, the magic bullet may fail late on in the procedure when you’re really testing it on a lot of humans and you find that it’s just not effective, or worse, has deleterious impacts. And at that point, you would have to abandon the drug. And that’s a late failure. And at that stage, you’ve spent many millions, if not billions of dollars on the development, and you may have taken 10 years.

So it just fails very easily, on average, far less than let’s, say, 10% of the attempts that people make to design a drug actually make it through to market. It’s a very slow and sequential process, and it’s hidebound by regulatory procedures that have to be met at each and every stage imposed by things like the FDA.

IRA FLATOW: And how are researchers hoping to change all of that with AI?

PETER COVENEY: So what everyone is trying to do is to remove the lengthy and time-consuming, labor-intensive experimental efforts. So if you can come up with something that’s in silico, meaning on a computer, that makes reliable predictions before any experiments need to be undertaken, that should short circuit the time of developing the drug.

Now, the current hype is around artificial intelligence being able to do that. And that’s predicated on the idea that these techniques are extremely powerful and as applicable to each and every problem that we have to face in our world. So drug discovery is only one of those.

But the idea is that somehow armed with enough information and data on the disease case, the protein you’re after, and candidate molecules you can race through and test which ones are going to be most effective with great speed, using an AI system. It will have to have been trained with an awful lot of data that’s reliable in order to be used in the way most people think of it.

And after it’s been trained, you just present the input that you’re looking for, which of these candidate molecules will bind best to the target, and it should hopefully spit out the answer often in seconds at that point.

But what’s at risk here is an illusion of thinking that all we need to do is find the magic bullet at the start. And then we found the drug. That’s, in a way, only the beginning because you then have to subject that candidate molecule to lots of tests that end up also in the clinical trials stage, and the failures can occur late.

Now, AI is particularly good at doing searches over huge, we call them, libraries. And we might be talking about many billions of molecules that it can go through very quickly to see if it can find, what we call, hits, meaning the molecule looks like it’s got a good geometric fit inside an active site on a protein.

That part of the problem can certainly be done pretty effectively with AI, but that’s really only the very first step in the process. The reason for that is that we really need to have a good understanding of the mechanisms by which molecules interact with proteins. We can’t just find the answer by looking things up in an existing database.

IRA FLATOW: Very interesting. Let’s talk about some of the businesses of these companies. Can you tell us, for example, what happened to a company called BenevolentAI? It seemed like this was a big name in the space.

PETER COVENEY: Right. So back in the mid-teens, we had the first wave of AI. And we were told that AI was going to solve the drug discovery problem. But in the end, we’re 10 years or more since BenevolentAI made its original claims. So that’s the time frame over which new drugs should be discovered. And we have nothing actually on the market that can be attributed in any way directly to AI.

So something has gone wrong. It’s overblown expectations predicated on things that I was saying earlier that AI is certainly good at and can accelerate some parts of drug discovery. But itself, it’s not the single thing because no single part of that workflow can be managed only by AI. It is one part. So that’s the limitation.

And that’s why BenevolentAI went from a massive multi-billion dollar rating after its initial flotation, and then it became almost valueless and has been taken over by another company since around 2023-ish.

So that approach didn’t work. The question you’re getting at implicitly there is why should we suddenly think that, OK, that was the old form of AI and now we’ve got this new one? It’s called generative AI, and that’s much better.

The point about generative AI, as its name suggests, is that somehow it’s generating new things in a way that somehow old-fashioned AI and machine learning didn’t. It’s a more sophisticated way of using data that it’s been trained on to try to make predictions that might be genuinely new.

And the generative algorithms that exist are now creating new molecules, quite possibly ones that have never been made before. And that gives us the opportunity to invent something totally new.

And that could be where the magic bullet resides. That’s the hope and, as it were, expectation of the new belief in the use of AI in drug discovery. But remember still what I said earlier. Even after doing that, I’ve got to take the molecule through multiple stages, optimizing it, testing it in laboratories, and then eventually to clinical trials. And I haven’t really done much in these discussions so far to claim that AI is going to help with a lot of those other parts of the problem.

[MUSIC PLAYING]

IRA FLATOW: After the break, it might seem that using personal data to train AI might be helpful for treatments, but there are some drawbacks.

PETER COVENEY: The idea that you can use clinical trials coming from Western sources to reliably predict how to treat patients in India seems a flight of fantasy.

[MUSIC PLAYING]

IRA FLATOW: Google has a startup called Isomorphic Labs for developing AI-designed drugs, and their chief AI officer, Max Jaderberg, recently said on the StarTalk podcast that the entire drug development industry needs to change how clinical trials are done, that they take too long, that mouse models aren’t very predictive of human success. What do you think of that?

PETER COVENEY: Now, I think this is a really good issue and point that he’s raising there– the future of medicine as a whole, it’s part of what I would call building digital twins. Because what we want, if possible, is not only to have in silico developments of the drugs, but also to be able to do in silico clinical trials as well, instead of putting these things into animals.

If we have sufficiently reliable computer-based models of relevant parts of the human body, well, it seems sensible to many of us to use the simulation methodologies there with patient-specific data on humans rather than using animal models. Indeed, we know in numerous cases now that if we do do this, replacing animal models by human data, that in silico trials tend to be a more reliable guide to success in the future.

IRA FLATOW: Yeah, because animal models are not really that accurate, are they, at predicting success in people?

PETER COVENEY: No, they’re not. And sometimes they’re totally misleading. And what’s more, this is also true between humans. Because humans are very diverse. And their own responses to a given drug can be completely different. So that’s in the end why you want a clinical trial.

Having worked up this magic bullet, you try it out on many people. If it’s a successful one-size-fits-all drug, typically it only really works well for about 50% of the population, and you hope the other 50% don’t get deleterious effects caused by them.

But certainly in areas like cancer, many of the drugs that do work work on far fewer than 50%. It can be 10% or less of the population because they depend specifically on something that might relate to the genome of those individuals.

So what you really want is some in silico representation of a group of people in a clinical trial that you can actually test these drugs on virtually initially because that not only eradicates animal testing, but in the first stage, also, anyone who volunteered in this way would only be giving their own data and putting themselves at no risk at all.

IRA FLATOW: So where do you get all the data that the AI models need, all the data from. From people that you want to feed in to the AI?

PETER COVENEY: If we’re talking now about clinical trials and so on, what we rely on is typically, shall we say, in the Western world, all the clinical trials work that has been done over decades and so on provides databases that can then be used in the way you’re getting at.

The problem with those is they’re clearly very biased because they do tend still to the moment be dominated by white males, which shows that they’re not exactly reliable just say for women as a category. But if you go to other parts of the world– and indeed, I was in India for the whole of January of this year, working with some health care providers over there.

I discovered that in India they haven’t really cracked doing clinical trials themselves. So they’re clinicians and doctors still lean on Western data sets in order to predict things for their patients. India, you may realize, is an incredibly heterogeneous country. So the idea that you can use clinical trials coming from Western sources to reliably predict how to treat patients in India seems a flight of fantasy.

AI has a role there because the amount of data you can end up with on individuals is a lot too. And AI can be very powerful with population health issues too. But if you’re trying to treat the patient in front of you, you need to the details of that patient, not a statistical average given by an AI system.

IRA FLATOW: So let’s talk about the path forward. If you had a magic wand, if I could give you a blank check to do everything you’d like to do, what is the path forward you would see for AI helping in drug discovery?

PETER COVENEY: I would see its role along the lines we said earlier. It can contribute, and I say it already does. It’s not fulfilling the bold claims that it’s discovering drugs right, left, and center or reducing time to solution to a couple of years from 10. Possibly it’s shaving off the time to solution from 10 years to 9. There are some arguments in that case. So it’s generic capabilities are useful throughout those workflows.

So there’s two things here. It’s a bit like weather forecasting versus predicting where the climate is going. Weather forecasting is good over a short period of time. Climate is about decades to centuries.

And in the human body, there’s the need to take action under some often emergency or short-term conditions. That’s like weather forecasting. And the wellness thing is predicting into the longer term. That’s more challenging. But that’s the way we expect things to evolve in the long term.

IRA FLATOW: Seems like a lot to bite off and chew here.

PETER COVENEY: Yes, there’s a huge amount that is there to be done. We’re not going to just suddenly solve this with one technique. I would say that the challenge here is we haven’t said much about what doctors and clinicians think of all of this, and it is a radical transformation in how we approach medicine, which isn’t really gaining traction in most medical schools yet.

But I do notice that the students who are doing medicine today are really trying to push their profession more in the direction of using predictive software, AI tools, coding, and trying to model things. So I think for the longer term, this will gain momentum.

IRA FLATOW: Well, we will keep watching it with you, Dr. Coveney. Thank you for taking time to be with us today.

PETER COVENEY: You’re most welcome.

IRA FLATOW: Dr. Peter Coveney is the head of the Center for Computational Chemistry at University College of London. He’s also the co-author of the book Virtual You– How Building Your Digital Twin Will Revolutionize Medicine and Change Your Life.

Thanks for listening, and don’t forget to rate and review the podcast, but only if you like the show, really. And if you have a comment or question or a story idea, our listener line, it’s always open. Call 877-4-SCIFRI. 877, the number 4, SCIFRI. This episode was produced by Dee Peterschmidt. See you next time. I’m Ira Flatow.

[MUSIC PLAYING]

Copyright © 2025 Science Friday Initiative. All rights reserved. Science Friday transcripts are produced on a tight deadline by 3Play Media. Fidelity to the original aired/published audio or video file might vary, and text might be updated or amended in the future. For the authoritative record of Science Friday’s programming, please visit the original aired/published recording. For terms of use and more information, visit our policies pages at http://www.sciencefriday.com/about/policies/