Oxford Sparks

Emily Elias: AI and big data has the potential to personalise medicine, and scientists are working on computer models of a brain – and not just a brain, your brain. On this episode of the Oxford Sparks Big Questions podcast, we’re asking: what is a digital brain twin?

[music]

Emily: Hello, I’m Emily Elias, and this is the show where we seek out the brightest minds at the University of Oxford, and we ask them the big questions. And for this one, we have found a researcher who loves to see pictures of brains.

Dr Andrea Luppi: Right. So I’m Dr Andrea Luppi. I am an early career researcher at the University of Oxford and the University of Cambridge. And as a job, I am a computational neuroscientist, which means that I look at brain data from humans and other species. I try to make sense of them, and I try to build models of how brains work.

Emily: Okay, so today we are talking about digital brain twins. These three words, I understand them individually. You put them together, I’m a bit lost. So walk me through, what is a digital brain twin?

Andrea: Right. So you can imagine a situation where a patient might want to have a particular problem that they have solved, and they might go to a physician. Now, what you want to be able to do is to give them personalised advice. We want to solve that patient’s problem, not someone else’s problem. And one good way that we can do that is usually trial and error, meaning that we might recommend some option that works for the majority of people. And if that doesn’t work for this particular patient, we recommend something else that works for maybe another smaller population. Trial and error eventually, it usually works, but we might want to get there faster.

So imagine if we could have a model, a simulation, a very simplified, admittedly, simulation of how someone’s brain works so that we could say, maybe based on a brain scan that we have of them, this might be the treatment that looks like it might be better. There’s a lot to unpack there, of course, but the idea is if we can simulate brains on a computer, at least some of the key ways of how that brain works, then we might be better able – and this is the digital twin part – to make personalised recommendations.

Emily: So how would you get all my brain information then? Where is it coming from?

Andrea: The way we usually do it is with functional MRI scans, which means it’s non-invasive. You just lie in the scanner for ten, maybe fifteen minutes. You don’t have to do anything. And we look at how different parts of your brain activate more or less at different points in time, in sync or out of sync with each other. And then we also know which parts of your brain are more or less physically connected to each other.

So what we can do is we put that information into our model, into our computer, and then we get a simulation of how those different parts of your brain might interact with each other to produce something that resembles the patterns of activity that come out of the brain scan. It’s very simplified, of course, so there’s a lot of detail that we will be missing out. But the point is not to capture everything. That would definitely not fit in a computer. The point is to capture enough that it can tell us something useful about what’s going on in your brain.

Emily: So if I had a disease and you had scans of my brain, I could say, “Hey, will drug A work best for me, or will drug B work best for me?” And you’d be able to go, “Do, do, do, do, do,” the machine would say, “Let’s go with drug A.” Is that kind of the idea?

Andrea: That’s right. That’s the long-term goal. We’re not quite there yet, but we are making progress towards that. But that is very much where we want to end up.

Emily: And how much data do you need to make a twin of my brain? Because my brain, it’s pretty crazy. Like, there’s lots going on in there.

Andrea: Yep. And right now we can’t read minds. So your brain would not fit on any of the computers that we have. So we’re using a very, very simplified readout of what we’re getting. And the whole point is, is that informative enough? So we’re reducing a lot the size. And in a sense, we’re tweaking it a bit like a fingerprint. Your fingerprint does not contain the entirety of you, but it’s useful for us to be able to tell, “Hey, that’s you, not someone else,” and act based on that.

Emily: So my brain and your brain are obviously two different brains. So they would almost look in the models like a fingerprint, where one type of drug or something might react to your brain differently than my brain.

Andrea: Exactly, right. So this is called, in fact, brain fingerprinting. People have already been doing it for maybe about a decade. And the point is, right now, we can do it with real brains. If you have two brain scans of two different people, you can tell which one comes from which person.

The problem right now is that until very recently, the kind of brain models that we could make did not have that property. So my model and your model would have been very difficult to tell apart, which is not great when what you’re trying to do is a digital twin rather than a digital second cousin once removed. But now we’re getting a lot closer to having models that are more personalised. And that means that in the future – and that future has come a lot closer now – we might be able to make personalised recommendations.

Emily: So where are we at right now?

Andrea: We’re at the point where we can now make brain models that, although simplified –and I know I keep repeating that, but it keeps being true, and it’s worth bearing in mind – we can actually tell apart which model comes from which person. And we can predict a little bit what their brain activity is going to look like.

So we’re starting to have the tools that allow us to say, given these two potential treatments, which one might we be able to implement? The next step, of course, is to show that the prediction itself works. We don’t just want models that tell me apart from you. We want models that tell whether drug A or drug B is going to work better for me or for you. That’s the next step.

Emily: And so what kinds of diseases are we looking at treating with this sort of thing?

Andrea: Yes. Well, of course it would have to be diseases that have something to do with the brain because we’re talking about brain models. That should go without saying, but I’ll say it anyway.

And then there’s a couple of different considerations that we might want to look at. For example, brain scans can be expensive, and the modelling itself does take still labour. So probably it would have to be the kind of diseases or conditions where a brain scan would be acquired anyway. So it’s not going to be, at least for the first time, any run-of-the-mill problem. It’s going to be something more major that someone might need to already have had a brain scan.

Examples can be if you need surgery for a tumour. Surgeons need to know where the brain looks like so they know where they’re going in. Or if you are, unfortunately, a patient with a disorder of consciousness like coma. Those patients are also routinely scanned, meaning that we already have all the data that we need. And if we have good models, we could plug those data into the model.

Emily: Where are we at in timelines in terms of actually seeing that happen? Because that feels like the future to me.

Andrea: Yeah. So my hope is that this is achievable maybe within the next five years or so. Myself and colleagues as well are working towards making it achievable, towards getting the right kind of data, the right kind of simulations to really scale this up.

Because right now, of course, it wouldn’t be feasible for even just a single hospital to send their data to us and allow us to do that. But if we can make a platform that is relatively easy to use, even for very busy clinicians, then there could be some of the next steps. So first is showing that it works, showing that it works at scale, and then making it usable for others, and by others, at scale.

Emily: And how do you make sure that what you’re getting is reliable information? Because obviously when you’re dealing with somebody’s brain, say, like in a cancer situation where they’re trying to make sure the margins are very clear so they have a good chance of survival, how do you make sure you got the right information to the doctors?

Andrea: Yeah, there’s a lot of different ways that you might want, lots of different things that you might want to look at. Right now the main thing is, are we simulating the condition well, and are we simulating the treatment well? And those are the very next steps that we’re going to be trying to get.

So when you have someone’s brain scan, you can check, am I getting the correct brain scan for that person? If I look at that same person’s scan maybe a day later, does my brain model correctly recognise the person? And then can I tell apart the brain scans of the people who are healthy from brain scans of the people who have the same disease? If we can’t, then there’s something wrong with the model. If we can, then we can move on to the next step, which is trying to find what makes the disorder model look more like the healthy model.

Emily: Timelines of this. Where are we at with your research in terms of actually seeing something being tested out in a clinical setting?

Andrea: I think clinical setting is going to take a little longer. Right now, most of the applications are about making sure that what we’re doing is actually feasible and actually gives us reasonable results. And for that, there are existing data sets that we can use because that kind of data collection takes time.

Fortunately, there’s not huge numbers of people who have the kinds of conditions that might need this on a more urgent basis, meaning that it takes time to collect those data. Some of them we have available, and so these are the ones that we’re working on. Once we have shown that we can reach effects that were actually observed for real – for example, we might have data before and after a tumour surgery, and we might ask, can the model correctly predict the effect of the actual surgery that happened just by looking at the data before the surgery?

Once we are in that situation, then we might try to propose the use for clinical application. But of course there are good regulations in place. Those regulations are there for good reasons, so that will take a bit longer for sure.

Emily: Obviously you are dealing with a lot of information. How big do these servers need to be that are storing all of this detail about our brain in order to make these calculations? Surely there has got to be a ton.

Andrea: Yes. So the size is not trivial. We’re talking terabytes rather than gigabytes. But it is also because the models are very good at compressing what really matters that it is still feasible. And so we’re still managing to do that. But indeed, we live in the age of big data, and this is nothing if not a big data exercise.

Emily: What makes you excited to work on this project every morning?

Andrea: I think it goes mainly to two things. One is really the basic science of it, the feeling that we’re getting closer to understand, at least in part, how brains work, and being able to say these regions are connecting in this way and interacting in this way, and because of this then that person is healthy and that person is not so healthy.

I think this is just a very exciting time that we have the capacity to even get that kind of understanding. My predecessors, who were working even just fifty years ago, one hundred years ago, they could not have dreamt of getting to this point. And this is just extremely exciting. And the kind of tools that we have, the kind of data-sharing initiatives that we have, a lot of that we owe to them.

This is a very exciting moment to be working on this question. We feel like we’re on the cusp of making big discoveries. And the other, of course, is that these discoveries are not happening in a vacuum, right? They are happening with data from real people, and we might be able to use them to help people.

We’re not in a situation where there’s only two centres in the entire world that have the kind of computers that this requires. We’re in a situation where this is actually usable by many. If we get our models to work, we could get them to patients within five, ten years, within my lifetime. That is very exciting.

[music]

Emily: This podcast was brought to you by Oxford Sparks from the University of Oxford, with music by John Lyons. And a special thanks to Dr Andrea Luppi.

Tell us what you think about this podcast. Please get in touch. We are on the internet at Oxford Sparks, or go to our website, oxfordsparks.ox.ac.uk

I’m Emily Elias. Bye for now.

[music]