AI identifies gene interactions to speed up search for treatment targets

Benjamin Thompson

Welcome back to the Nature Podcast. This week, the AI that identifies gene interactions...

Nick Petrić Howe

...and assessing the impact of rocket noise on wildlife. I'm Nick Petrić Howe...

Benjamin Thompson

...and I'm Benjamin Thompson.

<Music>

Benjamin Thompson

This week, Nature has a paper by a team of scientists who have developed a way to use an artificial intelligence system to predict the interaction between genes in human cells. They hope that it will speed up the mapping of gene networks and have already used it to identify genes that are potential therapeutic targets for a type of heart disease. Reporter Alex Lathbridge takes up the story.

Alex Lathbridge

Going on inside your cells right now are a countless number of biological processes that are essential to things like the development and maintenance of healthy tissues. These processes are regulated by a complex system of genes. In many cases, right now at least, researchers don't know how they're all connected. And figuring out these networks is no easy task. As Christina Theodoris from the Gladstone Institute of Cardiovascular Disease in the US explains.

Christina Theodoris

Let's say there was only two genes in any cell and you were only looking at those two changes, then you could pretty easily see what happens when one goes up and the other goes down. But when you start looking at 20,000 genes that are within the genome, that's when it becomes much more complicated. Because many of these genes are wired in a way that they affect downstream genes by increasing or decreasing their expression. Those connections can change in different states across time and development or space within the body in different tissues. They can also be rewired in disease states. And all of these many genes interacting in a coordinated way becomes a very complex process that's difficult to untangle, which is why we've turned to artificial intelligence and machine learning to help us.

Alex Lathbridge

This week, Christina and her colleagues have published details of an artificial intelligence system called Geneformer that they hope will speed up the process of identifying how genes interact. But to allow Geneformer to learn the fundamentals of how genes interact within networks, it had to be trained on a huge dataset, in this case, publicly available data on the transcriptomes taken from 30 million cells.

Christina Theodoris

Transcriptomes, or datasets, where we've measured the expression level of each gene in the genome, which tells us how active or inactive the gene is. And these datasets are measured in individual cells from different cell types, developmental stages or disease states. And so by collating all of these individual cells into a large-scale dataset, we can then train the model to understand how these gene networks change to control the state of the cells in each of the different settings. The model through a completely self-supervised way was able to learn to pay attention to genes that were really central within gene regulatory networks. And without having previously been told that these are transition factors or central regulators within a network, it was able to learn that on its own.

Alex Lathbridge

And the team could use Geneformer to virtually ask questions of simulated gene networks.

Christina Theodoris

We can take cells, take the set of genes that are expressed there and delete one of them and then ask what does that do to the gene network within the cell? And what does that do to the cell state? So, for example, if we take a healthy cell and we delete a gene, does it move to a particular disease state, or vice versa, which we refer to as in silico treatment analysis, can we take a disease cell, delete a gene and see whether that moves it back towards a healthy state, which would indicate that that gene could be a candidate target for therapeutic intervention.

Alex Lathbridge

Central to Geneformer's function is a process called transfer learning where the AI can apply information, it's learned about one situation to another one, think of it like learning to drive a car, you train enough and you can drive it, then without too much difficulty, you can transfer that learning to drive a van. So one of the team's key objectives was to see how Geneformer's learning based on a huge dataset showing the gene activity inside cells could be applied to situations where activity data is sparse. And where this in silico treatment analysis —predicting how removing a gene could affect a diseased cell— could have real impact. And so their focus turned to cardiomyopathy, a group of diseases that affect the heart muscle.

Christina Theodoris

And the reason we were interested in studying cardiomyopathy specifically is because it is a progressive disorder. So we would have the opportunity to intervene and prevent the progression if we had a targeted medical therapy. And this can make huge impact in the lives of these patients that often progresses to heart transplant when their heart fails. And because it's a relatively rare disorder and affects the heart tissue, which is hard to sample in the clinic, there are limited datasets specific to this disease. So when we apply the approach of transfer learning, we were able to predict candidate genes whose targeting would improve the disease, even though we only have limited data specific to cardiomyopathy. And in doing that, in silico, we're able to do so much more efficiently, both with regard to time and cost to really come down to a prioritized list of predictions that we could then test in the lab more easily.

Alex Lathbridge

Geneformer came up with a list of priority genes. Ones, which it predicted will improve the state of the cell if removed from the network. In the lab, the team set about testing heart muscle cells, known as cardiomyocytes to see if inactivating the genes could make a difference. And they found something,

Christina Theodoris

We were able to identify candidate therapeutic targets that when we tested them experimentally in the lab, did indeed improve the function of contracting of these heart muscle cells. Additionally, we were able to identify a gene that was not previously known to be key for cardiomyocytes were when we knocked it down, it disrupted their ability to contract. So we were able to use this transfer learning approach to make new discoveries and biological insights.

Alex Lathbridge

While there is some transcriptome data specific to cardiomyopathy, this has come from cells taken in heart biopsies, generally when patients are at a later stage of the disease. So one of the goals of the lab is to create a model that can help identify the potentially important therapeutic targets at earlier stages of the disease, allowing intervention to make a bigger impact. And while Geneformer's, predictive power is promising, Christina and the team know that when it comes to mapping complex networks, things are complicated.

Christina Theodoris

When you are testing these changes in the model, you are really only modeling that first step of the change. So removing that gene, does it shift to the cell towards the state of interest. So you're looking at the direction of the shift. In reality, if you were to remove that gene, there'd be many downstream consequences that would occur over time, that would change the cell more and more, until it became to that downstream final state.

Alex Lathbridge

To address this, the team are working to develop a model that could help build a better picture of how the network progresses after those first changes are made. But despite some of its current limitations, Christina and the team think that the approach that they've taken with Geneformer could be a powerful tool for researchers more broadly.

Christina Theodoris

So I think one thing we want to emphasize is that by seeing a very broad range of human tissues, the model is able to improve predictions in many different applications and different settings, not just cardiovascular disease that we particularly were interested in. And there's a lot of potential and researchers using this to answer questions in their fields and diseases of interest.

Benjamin Thompson

Christina Theodoris there. For more on this story, head over to the show notes for a link to the paper.

Nick Petrić Howe

Coming up, how researchers are going to figure out the effect of rocket noise on wildlife. Right now though, it's time for the Research Highlights with Dan Fox.

<Music>

Dan Fox

Researchers have found and identified bacteria that could help break down some of the 'forever chemicals' accumulating around the world. PFAS Per- and polyfluoroalkyl substances are a group of widely used synthetic compounds that have been nicknamed forever chemicals because of their incredible persistence in the environment. This longevity comes because PFAS molecules are full of fluorine atoms, which binds tightly with carbon, forming bonds that can't be easily broken down by microbes in the environment. But chlorine-carbon bonds are weaker. So the team behind this new research wondered if the chlorine atoms in some forever chemicals might give microbes a point of attack. The researchers obtained microbe-rich sludge from a local sewage-treatment plant and exposed it to a variety of chlorine-bearing PFAS. They found that, after the bacteria first removed the chlorine, they could then remove almost 80% of the fluorine atoms. Finally, they used genomic sequencing to identify the species of microbes most closely related to those in the sludge. The team hope that their findings will lead to new chemical cleanup technologies, and perhaps the creation of biodegradable versions of forever chemicals. You can read that research in full in Nature Water.

<Music>

Dan Fox

When it comes to gummy sweets, texture is just as important as flavour. Now, scientists have used some clever chemistry to identify the precise combination of ingredients needed to ensure that gummies keep their chew. If gummy sweets are stored for too long or in the wrong conditions, they can become harder and less satisfying to chew. In an effort to maximize how long they stay fresh and soft, researchers tweaked the factors that most effects texture: the recipe's concentrations of starch and gelatin, and the ratio of glucose syrup to sucrose. The team measured the properties of the resulting sweets during storage for different lengths of time, and at a range of temperatures. They then used the sweet's mechanical properties to estimate how ageing affects a crucial feature: the average distance between its molecules. Sweets with longer intermolecular distances are softer. They also estimated freshness by measuring other physical characteristics, such as water content, and pH. Out of the eight recipes studied, the team found the three resulted in optimal shelf life, including one of them, which surprisingly involved no starch at all. You can chew over that research in Physics of Fluids.

<Music>

Nick Petrić Howe

Finally, on the show, it's time for the Briefing Chat, where we discuss a couple of articles that been highlighted in the Nature Briefing. Ben, what have you got for us to chat about this week?

Benjamin Thompson

Well, I've got a story that I read about in Nature. And it's about some research that's helping to explain why chronic stress can inflame the gut. And it could have important consequences in how this inflammation is treated.

Nick Petrić Howe

Well, I think I've heard things like this before. So maybe give me a little bit of a refresher. How does stress affect the gut?

Benjamin Thompson

Yeah, it's true that psychological stress is known to worsen gut inflammation, particularly that caused by certain bowel diseases like inflammatory bowel disease, and quite why this is hasn't really been well understood. And these conditions can be very serious, right? Inflammatory bowel disease has a bunch of symptoms: abdominal pain, diarrhoea, and fatigue to name just three. And the two main types of IBD are ulcerative colitis, and Crohn’s disease, okay. And this can be debilitating for people who have them. And it's known that stressful events often precede IBD flare-ups, okay, there's an association between the two. But really why that is, hasn't been particularly well understood. And it seems like researchers have found a link between the brain and the guts.

Nick Petrić Howe

So there's a link there, but we're not quite sure how it works. But does this new study tell us a little bit more about what exactly is going on here?

Benjamin Thompson

Yeah, they have an idea. And they've studied it in mice. And these mice have IBD-like symptoms. And this pathway, as I say, starts in the brain and ends in immune cells in the gut, in particular. And so what they found is that after a surge of stress, the brain sends signals to the adrenal glands, and these release chemicals called glucocorticoids. And these are absolutely key to this whole story, okay. And these are released to the rest of the body.

Nick Petrić Howe

Right? Okay. So how are the researchers trying to figure out what these glucocorticoids are doing?

Benjamin Thompson

Well, initially, the article said that the researchers thought that these glucocorticoids were directly affecting immune cells in the gut, right. But actually, they showed there's a little bit more to it. So actually, what's happening is these molecules are making their way into the gut, and they're acting on neurons in the gut, and in particular, a specialized type of neuronal cell called a glial cell. And when activated, these themselves activate the immune cells. Okay, so we get into this sort of end-game now. And the immune cells released the special chemicals that are usually designed to ward off pathogens, but in this case, there are no pathogens present. So what it does is it leads to the inflammation in the gut, which is a sort of key symptom of IBD. And that can cause a great deal of discomfort for people.

Nick Petrić Howe

Now understanding this pathway a bit better, does that leave the door open for sort of new therapeutics and things like that?

Benjamin Thompson

Well, actually, it leads to a bit of a bit of a paradox, an unexpected paradox. So one of the treatments that's used for it IBD is glucocorticoids.

Nick Petrić Howe

Oh.

Benjamin Thompson

Okay. Yeah, I know. So this is something that the researcher is trying to figure out to be honest with you. And they say in the article, this came across as a big surprise. And what maybe seems to be the case, but sort of TBC, I guess, is that it's about timing. So when someone's treated with glucocorticoids, that's quite a short-lived treatment. And in a short time, it seems like these molecules might have an anti-inflammatory effect. But when it comes to things like stress, and this kind of chronic levels of stress that are sometimes associated with IBD, this longer timeframe seems to flip things on its head and turn this into a pro-inflammatory situation. But as I say, this is currently in mice. And there's lots to understand.

Nick Petrić Howe

So there is this sort of paradox, then where does that leave the research? What are the sort of next steps now?

Benjamin Thompson

Yeah, well, it shows something that I think is important that there's more evidence of the brain's ability to drive inflammation in, kind of, distant parts of the body. And I think we've covered on the show how, you know, brain and body are so interconnected when it comes to disease states. And and they say that perhaps in future this research means that the treatment will lead to combine drugs, maybe with stress-management techniques, and that could be more effective than just drugs, and potentially know more about this pathway could be new therapeutic targets that go for just it. And there's also evidence that stress can cause inflammation in potentially other parts of the body as well. So maybe by knowing a bit more about it in this situation, it could offer new insights into how the brain and body are connected in situations like that.

Nick Petrić Howe

Well, obviously, there's lots more to understand here. But hopefully, in the future, we'll hear more about it possibly new therapeutics for people who have these conditions. For my story this week, I've been reading in Nature, about how noisy rockets are, and have researchers are trying to understand how the noise in that launches is affecting wildlife.

Benjamin Thompson

Well, I spoke to a researcher a couple of weeks ago, who was at the launch of the JWST in French Guyana. And they said that there was this impossibly loud rumble for a bit as the rocket took off, is this the sort of place they were looking at?

Nick Petrić Howe

Yeah, I mean, this article focuses on a couple of launch sites in the US. So there's one in Texas, one in Florida, and one in California. And they're all in biodiversity hotspots within the US. So they're far away from people and obviously close to wildlife. So there could be a big impact. And as you said, as well, like, they're very noisy. So, for example, the launch of Artemis I, which you probably remember, was recorded to be 127 to 136 decibels, which is louder than a clap of thunder from several kilometers away. So it's pretty loud. And so it's understandable to think that there may be effects on wildlife. But the reality is, we don't necessarily know what that is.

Benjamin Thompson

So what sort of things are researchers wanting to know?

Nick Petrić Howe

Well they want to know if such noise is having an effect on animals' behavior, whether it's causing them stress, or whether it's affecting like things like birdsong. For example, if you and I hear a loud noise, we may shout to compensate, do birds do a similar thing with their song? And would that have an effect on them? So they want to understand what the effects are. And this is especially pertinent now, because the number of rocket launches is increasing substantially. So for example, in this place in California, where a lot of rockets are launched called Vandenberg, it historically had about 5 to 15 rockets launching per year, but by 2030, that's going to go up to 50 to 100 per year. So there's a substantial increase in the number of rockets that are going up, and that may well be having an effect on the wildlife. And we need to understand that.

Benjamin Thompson

And how best to sort of go about looking at what that effect might be, then?

Nick Petrić Howe

Well, the researchers are doing a three-year study where they're using cameras and microphones to assess the impact on the behavior, so things like the birdsong and how the animals are reacting. Do they look, I guess, startled, or what's their sort of reaction to these very loud noises? And to have some sort of long-term monitoring to see if that has any impact on the species around, especially some of the endangered or protected species that exist around these sort of launch sites.

Benjamin Thompson

And do they suggest any things that could be done to maybe mitigate these effects? Obviously, we don't know what they're going to find yet. But But what are they put forward?

Nick Petrić Howe

So they've got funding for three years, but they're actually going to try and do this for more than a decade. But as it stands, rockets already have an amount of sound dampening that goes on. So if you've ever watched a rocket launch, you may be familiar with the big fire trenches that are underneath when a rocket goes off, and this sort of channels where the rocket blast goes, and this also helps to sort of dampen the noise. And also for many lunches, they use a lot of water underneath the rocket to absorb a lot of noise. So things like this could be done or changed. Or the other thing is, they may identify sensitive periods when the wildlife are particularly susceptible to a lot of noise, maybe it's their breeding season or something like that. And that could inform whether rocket launches should be sheduled at a different time like to avoid this sort of very sensitive period for them. So it'll be interesting to see how this research shapes up going forward and what that may mean for how we launch our increasing number of rockets.

Benjamin Thompson

I mean, I suppose we do talk about rocket launches a lot on this podcast, but really, we're thinking about what happens after that has happened and and I think, really thinking about what's going on at the time, as well as is an important avenue that I'm maybe hadn't fully considered. So yeah, very interesting to see where this one goes. But let's leave it there for the Briefing Chat this week, and listeners for more on these stories, and where you can sign up to the Nature Briefing to get even more like them delivered directly to your inbox, check out the show notes for some links.

Nick Petrić Howe

And that's it for this week. Just before we go don't forget you can keep in touch with us on Twitter, we're @naturepodcast, or you can send us an email to podcast@nature.com I'm Nick Petrić Howe...

Benjamin Thompson

...and I'm Benjamin Thompson. Thanks for listening.