Digital Twins in Medicine: The Future of Clinical Trials
An episode on digital twins, the virtual replicas that began in NASA's space program and are now being applied to medicine. Aaron Smith, founder and machine learning scientist at UnLearn, explains how his company builds models of patients' likely disease progression and uses them to run clinical trials with fewer subjects. He also discusses rare diseases, regulatory requirements and partnerships with pharmaceutical companies.
Transcript
Welcome to Agents of Tech, the podcast that dives into the transformative innovations and key ideas driving the future. I'm Autria Godfrey, joining you from Washington, DC.
And I'm Stephen Horn in London. Today we're talking about digital twins, a concept that's not just futuristic, but outright mind bending, at least for me, Autria. Imagine your car, your city, or even your own body having a virtual doppelganger, constantly learning, adapting, and simulating possibilities to make life smarter and more efficient.
It's truly amazing how this started with aerospace, and now we're applying it to health care and research. And it's not science fiction anymore. It's actually science reality. And it is certainly, as you mentioned, changing how we plan, predict and problem solve across multiple industries. So let's dive right in and unravel the twin truths of this digital revolution.
The concept of a digital twin was first adopted by the NASA space program in the 1960s, to simulate a spacecraft and be able to debug flight issues in real time as they arose. This concept was successfully used during the Apollo 13 mission, when the spacecraft suffered a malfunction, and the NASA team had to simulate conditions aboard Apollo 13 to bring back the spacecraft and the astronauts safely back to Earth.
It sounds like a like a sci fi movie, but it's reality. It's incredible.
Digital twins began as static replicas, a very straightforward digital version of a physical system. But of course, as innovation marched forward, so did the twins. We saw the rise of the "mirror twin", the "shadow twin", and now the "intelligent twin". This cutting edge version is more than just a copy. It's a self adaptive, powered by artificial intelligence with the ability to learn reason, act autonomously, and even communicate with other twins.
So Autria, how do you know this isn't science fiction? Because, I mean, it sounds crazy to me that we have these digital twins that just kind of, who knows, have you got a digital twin?
I hope I have one.
It's crazy. I mean, we're talking about 1960s, we're talking about NASA, we're talking about science. And as you quite rightly say, it's stuff that's here today, right?
Right. And now with the rise of artificial intelligence, we're able to take, you know, our intelligent twin, and hopefully, as we will dive deep in this episode of Agents of Tech, figure out how it's being used in the medical realm and quite possibly helping treat and even prevent diseases and other metabolic issues.
We're actually going to be looking, Autria, at how, in this particular episode, how we look at clinical trials and how this approach and how this technology will help us. We'll be talking to Aaron Smith, founder and machine learning scientist at UnLearn, an AI company that's revolutionizing clinical trials.
Right now it's easy to see a future where, the types of acceleration that we're getting by applying AI in clinical trials and in this way will be applied in essentially all phase three clinical trials. And in aggregate, that's a pretty big accelerant. I feel like that's inevitable. But the question is how quickly can we get there?
Remember, if you are enjoying our show, please don't forget to subscribe to Agents of Tech. This really helps develop the show and shine a light on cutting edge research. So please like and thank you.
Well, Autria, I'm absolutely delighted as always to be joined by Laila Rizvi who as you know is the Director of Research and Strategy at WebsEdge the company behind the podcast. Laila, great to have you here.
Great to be here. I'm really excited about this episode.
Because this really ties into some of your own research and work, doesn't it?
Yeah. So my background is immunology, and I remember my professor standing up there and saying that it is impossible to simulate the immunological kind of interactions that you see with the immune system because of the various scales of it and the fact that it's just random and unpredictable. And physics based models like differential equations, don't really allow for biological systems to be, simulated. But with AI, that could completely change. I mean, the idea of of being able to predict uncertain systems and random systems is extraordinary, because the world's random that the world we live in is uncertain.
It's interesting, isn't it? Because with these digital twins, they actually evolve as we evolve. So it's not just, you know, it's not just a static picture of who we are. It sort of adapts and evolves and, you know, to the same factors that we adapt, evolve to. So how did we get here from the realms of science fiction? How do we get to the point where this is now a reality?
I think it's weird because it's incremental kind of gains and improvements. And then before you know it, you open your eyes and you're like, oh, this is quite extraordinary. But whilst you're doing it, it doesn't seem like it's too crazy. I mean, it's at the end of the day it's really kind of statistics. Right. And I think if you're not there at the incremental stages of improvement, then you think magic, right?
Still seems pretty magical to me. What about you Autria, is it magic to you?
It's absolutely magic to me. And kind of to Laila's point, I feel like it was a slow drip until it was a flood. You know, this has kind of been like, we've been working on this since the 1960s. And now all of a sudden we've got these digital twins that are being utilized across, you know, multiple different platforms and different industries. And, I mean, I just keep going back to how this is something that was originally, created and used, you know, like you said Stephen with NASA at Apollo 13 and, you know, the fact that we were able to take, okay, this actually worked, you know, to get these astronauts back to Earth safely, how can we take this and pivot and use this to what we're now doing with, you know, disease models and precision medicine and clinical trials? So I think it's really, really amazing the breadth of these digital twins.
And I think, Autria, your point on kind of that knowledge being used in the 1960s speaks to even the wider mission of WebsEdge, that we ultimately build on the knowledge of the previous generations. So it doesn't just kind of come out of thin air. So we currently think that we're the smartest that we've ever been, and there's no way anyone could have imagined the technology we have. But we've built on the discoveries and the ideas and the conjecture of the previous gen.
So we're going to speak now to Aaron Smith, founder and machine learning scientist at UnLearn, an AI company that is revolutionizing clinical trials. Aaron, appreciate your time today. Can you start small by just kind of telling us about UnLearn?
UnLearn is an AI machine learning company. If you watch us, you know that we primarily work in the clinical trial space. Our core technology centers around building machine learning models called digital twin generators. These are models that generate, likely outcomes, probabilistic outcomes, through time of patients who have particular diseases, for example, Alzheimer's, ALS and so forth. How their clinical measurements, labs, etc. are likely to evolve over time, given that they have that indication.
And then we apply this, in many different ways, but primarily we're focused on accelerating clinical trials. So we work with pharma sponsors developing new therapies. We help them design and execute clinical trials that have, protocols that incorporate digital twins into the trial protocol, that ultimately allow them to run the trials more efficiently. So run the trials with fewer subjects without, making any other trade offs in terms of power. Basically the ability to detect an effective drug in the trial so you can recruit faster, it's cheaper, and you can get to the finish line faster. So if you have an effective drug, it'll be on market more quickly, and you can do it for less money.
So, Aaron, why did you choose this particular route of looking at, clinical trials rather than say, patient care?
First of all, clinical trial data is very high quality. So for the most part, our models are built on clinical trial data. So data generated, taken from control arms of clinical trials, in some cases both arms. But it's a matter of what we have access to, registries. This data is typically collected for the purpose of trying to answer questions about how someone's likely to progress. So you can you can contrast that with something like EHR data, where really it's about informing the insurance companies what items, the clinical items they need to pay for. So, that's one reason we were focused initially on that kind of data.
Another it has to do with the use and, regulatory constraints. If you look at the risk profile of these applications, something like a personalized medicine application. So you use a prognostic model or some type of machine learning model to make a diagnostic device for an individual, that carries extreme risk. Which sets the bar extremely high for how accurate and reliable that model has to be.
Clinical trials, provide you a framework by which you can, make use of the power of these models without the same risk profile because you focus a little more on aggregate statistics of the population in a clinical trial. You want to know about how a population or a particular subcomponent of the population is likely to progress, under treatment and control without having to have the constraint that the model is 100% accurate for every subject in that individual clinical trial.
So one way to put it is this is a wedge strategy for us. We are excited about precision medicine and clinical decision support. But clinical trials is an initial way in the space, that I think is suitable to that type of data that exists right now.
You actually, by talking about the risk associated with it got exactly where I wanted to head next. And when you're talking about the aggregate, you know, of the entire population that you're, sampling, I guess. Is it less risky because you have more data to compensate for, say, an outlier? Yeah, I'll leave it at that. That's only one way to put it. That's probably a very dumbed down approach.
No that's certainly part of it. Like a clinical trial. The data that's collected will have outliers, right? The fact that one subject has some unusual response profile doesn't taint the trial. But it does go into to the statistical analysis and skews it a little bit. If enough subjects are sort of wonky in one way or another, you know, it would affect the statistics for sure.
So where do you see the biggest opportunities to transform clinical development then?
I think there's maybe two answers. One is maybe a little self-serving but shorter term, and then one is maybe in some sense long term and obvious. But is worth mentioning. So, you know, the shorter term example is, the way that we're approaching accelerating clinical trials, and we'll get to this later, is regulatory acceptable in all phases of clinical trials. The statistics of the approach guarantee that the primary concerns of regulators can be met. It's a matter of time that this type of approach will be applied across the board.
So right now it's easy to see a future where, the types of acceleration that we're getting by applying AI in clinical trials and in this way will be applied in essentially all phase three clinical trials. And in aggregate, that's a pretty big accelerant. I feel like that's inevitable. But the question is how quickly can we get there?
On the other hand, I think there's maybe a broader, answer to this question and it's maybe, a bit of a cop out, but there's a ton of super hard science problems that we need to solve in the drug discovery space. And if you think of developing drugs as both discovery and development, if the candidates that you're plugging into the development component, the clinical trial component, do have a really strong, prior belief that that that drug is going to be very effective and safe, you understand that safety profile ahead of time, then that puts a lot less burden on the clinical trial stage.
You could imagine a future where some of those key biological problems have been solved by virtue of new computational methods and, and so forth, such that clinical trials become more of a more perfunctory safety check, that that would be a dream. I mean, that would basically put everyone who's doing clinical trials out of business in a sense. But that's very much more long term. But, honestly, that is the best way to accelerate clinical development, is make sure your candidates are much better, much more targeted, much more likely to work.
And I now want to bring it kind of back to the details, I guess, in more your technical approach to this, to healthcare and to specifically placebo responses rather than individual outcomes. By focusing on placebo responses, how has that narrow focus informed your technical approach? Because you've got you've got a lot of people working on the application of AI in healthcare. But I imagine you require a very different technical approach when it comes to placebo groups.
At core, we build models typically of placebo response that you know, given a subject's baseline measurements, how are they likely to evolve over time? It's totally common for us to have "standard of care", quote unquote, you know, or control. In this case, the standard of care could have disease modifying medications. I think the issue that probably that you're getting at is whether we're building models of, treatment response to new treatments, because we, you know, we can certainly use data aggregated in our larger data sets that have whole clinical trials, both control arm and treatment arms.
From someone who's from a medical background as well. I think you automatically when you hear a placebo group, how are you modeling the psychological effect behind it. But I imagine your aim is to increase the statistical power of the overall trial, as opposed to trying to understand the why, but rather just getting to the what.
The primary way that we're applying digital twins in clinical trials involves running a randomized control trial and aggregating for each subject in the trial, at the end of the trial, not only the measured results of subjects on both arms, but also the predicted, digital twin outcomes of both sets.
Okay.
So treated and control. And you can do the statistical analysis in such a way that the prediction of someone's outcome under control helps reduce the variance of the treatment effect estimate in that trial. That turns out to be a very powerful approach. It's somewhat relies on the fact that we have a lot of data. You could try to do some clinical trial designs where you build a model that involves predicting a subject's response to a novel treatment. It goes, far beyond, what anyone's able to do at this point in most disease areas.
If you're interested in a systematic disease and you want to have some kind of prediction based on no actual like prior measured data about the interaction of this molecule, let's say in someone's body, you're interested in how they're likely to evolve. What people have done historically in the past is try to build kind of causal mechanistic models. We generally think that that is far beyond the scope of what anyone can achieve at this point, even if it were to be done with a machine learning approach, it would involve the collection of huge amounts of data that has never been collected before. Basically, experimental interaction.
We didn't think it was even in the cards, let's say, to try to solve that problem out of the gate. Whereas there's actually a ton of useful value that you can get from, historical clinical trial data, an indication to improve clinical trials. And in this way, if you consider especially indications where there have not been much progress, every Alzheimer's clinical trial makes a really careful scientific study of half of the population in that trial, they record tons and tons of data, this is a control group I'm talking about.
Effectively, the subjects between all those Alzheimer's trials are the same group of subjects, roughly speaking, there's some differences. But for the most part, the way trials are run is that all the knowledge that's gained from those studies is thrown out, never used in any clinical trial. You kind of leverage the strengths of clinical trials as they are now, rather than trying to do something which we effectively can't do with current tech.
And just moving on to, kind of the models you're using. I know that you recently rebuilt your entire machine learning stack in 2023. And I was wondering what kind of drove, that rebuilding and what were the limitations of the previous approach, which kind of made you want to rebuild the stack so you can do what you can do now? And what did that tell you about the needs of digital twin architecture?
We started off with a particular model architecture approach that, we developed quite early on. It was very, very rough and ready, and it worked pretty well. But from a modeling perspective, it had some warts. You know, just to be like, super detailed about it. And we didn't love the typical degree of bias that those models had. In many applications the bias actually doesn't matter, but we were interested in developing digital twin generators. So models that produce and essentially continue subjects records as if they were real. So we want to make those as indistinguishable from real subject records as possible.
And that's actually a super high bar. That's a much higher bar than what you would need in maybe any particular application that we have in mind, in particular clinical trials. You can do a number of different things and there's different constraints on the model that each application has. But we had this sort of global constraint of we want the model to generate subject records that are indistinguishable from real ones. So we put a pretty high bar on, in terms of the accuracy of those predictions. And so we did find an architecture that solved that problem more completely. So that was probably the number one sort of technical or scientific reason, let's say.
Another was that we, integrated a lot of newer, tools that had evolved in the ML AI space and ultimately that made it so that we could train models faster, we could experiment more quickly. The actual number one gain we got from doing this, refactor of our stack was that we can iterate more quickly, we can investigate new ideas, we can investigate new modeling approaches, swap in swap out components, and get to experimental results internally much more quickly. That turned out to be incredibly important. And still, make a number of changes to move faster in that regard as well.
Aaron, Stephen here again, when you're looking at rare disease trials, when you're looking with small population groups for those trials with rare diseases, how do you validate your model when you don't have as much data?
Yeah. You can't get something from nothing. So, you have to make trade offs. So I can give one straightforward example. Typically when we're blessed with a large data set, we can treat a big chunk of our data as training data. And then within that training set, we'll separate out some validation part and we'll do model training on the training part will evaluate the quality of the model on the validation part to make some decisions about how to do the modeling. But then we'll have this holdout test set that we use to completely hold out, to estimate or validate the model performance for the application.
In a rare disease, we might make the choice, we have made this choice in some cases, where we say, actually, we're going to use all the data trained on because we want the best model possible. And then we, in partnership with sponsors who have a proprietary trial, past trial in the disease area, for example, we say, you know, we'll do a partnership with you, we'll accelerate your new clinical trial in that area. But we will start off by estimating the performance of our model in this rare disease area with your data set as if, you know, as that is the held out data. So that's a simple example of how you have to make do with less data in some cases.
And I like to kind of pivot a little bit to like oversight and regulatory requirements because I know that you have, secured both EMA qualification and FDA alignment. So how did those regulatory requirements shape your approach to your model validation?
We built our approach to clinical trials from the ground up with the regulatory requirements in mind. We look at what regulators care about, what they're trying to regulate in phase two and three clinical trials and match the approach to that exactly. So there are a lot of companies that are talking about making digital twins these days. Sometimes they mean what we do oftentimes not the same thing, but in some cases, there's some parallels.
A lot of those companies are proposing to do things like synthetic controls where they fill in, basically add subjects to trial, which are synthetic subjects. That's fundamentally not what we're doing. And the reason we're not is because that doesn't control the type one error rate. So if you want to run a phase three trial and there aren't extenuating circumstances why you can't produce a control group, you're just not allowed to take that approach, no matter how much you really want to.
One way to think of our technology is you have these digital twin generators, and then you have technology which are statistical designs of clinical trials that incorporate predictions, from models as well as the measurements that are taken in a clinical trial. And the properties of those different, we think of them as adapters, are adapting to the digital twin model and the application, the properties of them are specified, are designed mathematically to control the things that regulators care about.
So what determines whether a disease is suitable for your approach or not?
It's wonderful to have a lot of access to a large quantity of high quality historical data that is nevertheless relevant to current clinical trials. If you had a disease where there's just one thing you can recognize which like completely describes how a subject is likely to respond, maybe a extremely prognostic biomarker. It's not quite as exciting as systematic diseases where there's tons and tons of measurements that all have a very small influence on the outcome. There's some real deep complexity there. Those are the most appealing examples to us.
But as we expand into new areas, the the most exciting prospect for us, I'd say in the near term is partnering with large pharma that have their own, basically franchises in an area where they have many, many clinical trials, worth of data, entirely under their own ownership, so inaccessible to anyone. And we can come in and partner with them and say, okay, we're going to do what we do. Build a full feature digital twin generator, as part of this engagement. And then we'll be able to use that for all of its applications, especially accelerating clinical trials in that area going forward with that partner.
Aaron thank you so much indeed for joining us. We really appreciate that. It's been such a pleasure. Thank you.
It was a pleasure.
Well, that was awesome. I thought it was just incredible that they're working on placebo responses, which I think is a completely different approach to what I've ever heard. Although we didn't touch upon it, they're addressing a huge issue where normally in a clinical trial, the placebo group doesn't benefit from the drug. Whereas now they can reduce the control kind of placebo groups so more people can have the beneficial drug if it works out, which I think is pretty awesome.
It's also interesting to me the way he said, this is the future, this is the way that all clinical trials are going to be going, you know, before too long. It's a really interesting use of data and machine learning and AI isn't it? So that was interesting to me.
This is all about statistics. It's just that if the statistical power of the trial is increased, it doesn't really matter how we get there, as long as we're reliably getting there. So it's in the same way as making kind of weather predictions. You don't need to know how the weather is being predicted as long as it's predicting it correctly, then you know you're good to go, which I think is pretty extraordinary. And, is going to be critical in medicine for kind of physicians and those running clinical trials to understand that.
You know, the digital twin for health care approach has a use and a place with both individualized patient care and also these clinical trials. Which to me, that's very encouraging. Really fascinating interview that does it for today's episode. If you enjoyed our discussion, don't forget to like and subscribe, because that helps us come up with content and improve the show. So until next time, thank you so much and goodbye.
Goodbye.
Until next time, goodbye.
Agents of Tech is brought to you by WebsEdge. We want to thank our studio engineers, Adam Deen and Cody Huynh our editor Abbie Harries, and our producer, Cath Sheehan.