Ep 117: The time of your life (with Rosemary Braun)
How should we study complex biological networks? How do cells keep time and stay in sync? What does it mean for a network to be resilient?
In this episode, we talk with Rosemary Braun, Associate Professor at Northwestern University in the Department of Molecular Biosciences and a member of the NSF-Simons Center for Quantitative Biology. Rosemary is broadly interested in learning whether “more is different” when it comes to complex molecular networks operating across different temporal and spatial scales. We talk with her about systems approaches to uncovering the “Rules of Life” and about circadian (daily) rhythms. She and her team use machine learning to understand emergent phenomena in networks, with the goal of helping medical professionals target treatments based on an individual patient’s circadian rhythm.
Cover photo: Keating Shahmehri
-
Marty Martin 0:07
Bonjour, mon ami! Comment ça va?
Art Woods 0:10
Nice of you to greet me in French while I'm here in France on a research trip. Honestly, though, your pronunciation kind of stinks.
Marty Martin 0:17
Ah, sorry, but I'm not surprised. It's been too many decades without much practice. I hope my high school French teacher isn't listening.
Art Woods 0:24
Moi aussi!
Marty Martin 0:25
Okay, seriously though, onto today's topic, the rules of life. In a sense, one could say we have lots of such rules already.
Art Woods 0:33
Yeah, like all living things are made of cells. All living things use DNA as a major means of inheritance.
Marty Martin 0:39
All living systems use homeostasis, and all living systems evolve over time.
Art Woods 0:43
But is that it? Are these really all the rules of life?
Marty Martin 0:47
Well, chatGPT says that's it, for whatever that's worth. But our guest today, Rosemary Braun disagrees. She thinks that there are probably many more rules that we're just now poised to discover.
Art Woods 0:58
Rosemary is an Associate Professor at Northwestern University in the Department of Molecular Biosciences, and she's a member of the NSF-Simons center for quantitative biology. Her expertise is in computational biology, and her goal is to reveal new rules of life.
Marty Martin 1:13
A few of her key questions include, what are the general circuit characteristics that govern how cells work? What can the structure of biological networks tell us about their functions? And how do our living processes coordinate in time?
Art Woods 1:25
Yep, all seem like big questions to me and worthy of the term "rule" if our team can figure them out.
Marty Martin 1:31
And what's helping them do that is the enormous growth of big omic datasets, as well as the analytical approaches that focus on whole systems.
Art Woods 1:40
Rosemary is using her background in physics and math to study how living systems are organized and self organized into a series of nested networks. She thinks that particular network configurations might underlie healthy and unhealthy physiological conditions.
Marty Martin 1:54
Think things like good sleep, wake cycles, and cancer.
Art Woods 1:58
We also discuss her super creative approaches to studying temporal changes in biological processes and several new tools that she and her team have developed for analyzing circadian rhythms and disease associated alleles.
Marty Martin 2:11
Ultimately, Rosemary wants to learn how living networks can be both robust enough to tolerate mistakes, but also flexible enough to adapt to environmental change. Surely, this kind of understanding would reveal new rules of life.
Art Woods 2:24
I'm Art Woods
Marty Martin 2:25
And I'm Marty Martin,
Art Woods 2:26
And this is Big Biology.
Marty Martin 2:42
Rosemary Braun, welcome to Big Biology. Thanks a lot for joining us to talk about networks and complex systems and other favorite topics of yours and ours. So I'm going to start simple with the motto on your lab's homepage, you say "our central goal," your central goal, "is to elucidate, predict, and modulate the behavior of complex molecular networks that underlie all living processes." So give us your favorite example of work in that area that's not from your lab, but fits that motto.
Rosemary Braun 3:12
Oh, wow
Art Woods 3:13
Or your lab if you really need to,
Rosemary Braun 3:15
Or your lab! We're going to talk about your lab for sure, but we thought we would start broadly.
Rosemary Braun 3:19
Yeah, no, this is such a great question. I think, as we look around the world, living systems are characterized by these beautiful self organizing processes. And that starts out at the atomic level, where proteins fold into macromolecular complexes that carry out specific functions within the cells. And then the cells need to interact with one another to give rise to tissue level organization. And that tissue level organization needs to be coordinated across the body. Organisms in an environment are acting in concert with one another. And so we see the self organizing dynamics at all scales. And, you know, when you ask me about my favorite work, I think, you know, the thing that comes immediately to mind is Phil Anderson's famous paper "More is Different." Right? And, you know, as somebody who was trained in complex systems research, this idea that when you put things together, they are more than the sum of their parts, is really compelling to me and kind of sets the philosophy for my lab.
Marty Martin 4:27
So can you say more about that paper? I think a lot of our listeners will know about it, but more is different as a grouping of words is... It's more is better, right? That's what we would usually say, but how is more different?
Rosemary Braun 4:42
What I would say is that the underlying ethos of complex systems research in general is this notion that you cannot explain the behavior of a whole system based on a reductionist view of what its individual parts are doing. And when you bring things together in a system, what you often find are these beautiful emergent dynamics that you couldn't have predicted, if you just knew about individual parts. I think when we think about modern biology today, the way in which this, you know, influences my outlook, and I think many people's outlook, is we have ways to probe living systems and incredibly fine detail, right? The expression level of each individual gene, whether or not a particular gene has a mutation, etc. What that means in the end for the way the organism responds to a particular perturbation, or whether it's susceptible to disease, in general, you can't predict that by looking at individual genes, you need to look at things as a whole. And I think that that's, you know, where we are now in biology today, we have the opportunity to start to do that.
Art Woods 5:50
Mhhm yeah, agreed. You used this term, "emergent properties" or "emergent phenotypes." And let's just dwell on that for a second. Because I feel like in some ways, that's a super interesting, maybe also a little bit of a fraught idea. So what do you think about emergent properties as an idea altogether? And then the related thing is about directions of causality? So is there something going on here where the larger system is exerting top down control of what goes on in the pieces that make it up?
Rosemary Braun 6:22
Yeah, the question of causality and which way the control comes in is one that's pretty important and pretty thorny to answer in biological systems, because they are adaptive, right? I was trained as a physicist, and, you know, in instances when you are looking at inert systems, you still see emergent behaviors, right? Things that you wouldn't be able to predict based on the behavior of their components. But in general, there isn't this two-way causality that the system responds to some environmental perturbation, by changing its component parts. In living systems, this happens all the time, right? There's some sort of environmental change or an environmental insult, and the system has to respond to that. And, you know, deciphering whether, when we're looking at data and we're analyzing it from a statistical perspective, deciphering whether what we are seeing statistically is a cause, or an effect of the phenotype that we're interested in, is still very much an open question. And unless you have real time series data, where you can probe causality, right? Or you run an experiment where you induce the perturbation, and you can definitively establish causality, it can be actually quite hard to determine what is cause and what is effect from just a snapshot of data.
Art Woods 7:47
Right, awesome. I guess a related question is about, just to circle back to the idea of emergent properties again. In some ways we define it as we don't understand what a system is doing, from understanding the actions and the identities of the individual parts. But how do you distinguish that from the fact of maybe we just don't know enough? And if we didn't know enough, and if we modeled them all correctly, the interactions between those parts, we would understand the behavior of the system?
Rosemary Braun 8:16
I think you've really hit on it, right? The fundamental question here is, what is the scale at which we should be probing these systems, right? Is it correct to be probing things at the single gene level? Or do we need to probe things at some higher level of organization, right? And how do we determine at what scale we should be probing these systems? You know, I like to talk about this question of how do you find the effective variables that a biological system cares about? What are they? Are they the behavior of a given gene? Are they the behavior of an entire gene regulatory network? I think we don't yet know the answer to that question. And I think that right now, where we are with the experimental technologies that have been developed, and computational tools that have been developed, what we can start to do is we can take the data that we're getting at this molecular, very fine grained level, and run our analyses at coarser and coarser scales, and maybe be able, through those analyses, to determine what scale we should be looking at the data. But I think it's an open question.
Marty Martin 9:29
Hmm I mean, this is really cool. Can we try to connect it back to the more is different that we started with? Is there an effort or there's some these special thresholds at which we have enough "more" that different happens? Or is there like too much more, where everything falls apart?
Rosemary Braun 9:49
So there are special thresholds at which something special happens. In physical systems, we see these as phase transitions where you know you go from the behavior that occurs at a very small scale, as you scale up, suddenly you have a shift in the behavior. Whether such things occur in biological systems, we don't know yet, right? And that's part of why I advocate this sort of multi- scale, can we take this data and analyze it across different scales, and see what we learn at each of the different scales, as we are analyzing the data.
Rosemary Braun 10:24
If I take gene expression data, for example, just to make things concrete, I can analyze it one gene at a time, right? And I end up with a list of genes that might be differentially expressed in the phenotype that I'm interested in. But then I can coarse grain, and I can say, okay, instead of looking at the expression level of every gene independently of one another, let me take that data and project it on to a gene regulatory network, and come out with a summary statistic that just describes how highly expressed the genes on this network are, right, without regard to their specific identities. And now I can ask the question, okay, for this gene regulatory network, do I see differences in my two phenotypes of interest, where one network might be highly active in one phenotype and not active in another? And so I can look at it at the network level. And then I can look at sets of networks, et cetera.
Marty Martin 11:20
Right, we're going to talk more about this idea in terms of the work you're doing now with circadian rhythms and others. So maybe we don't have to dig into it in detail right now, but what's your opinion on where the field now stands, i n terms of sort of single genes and expression networks of genes? The thing that seems to be percolating right now maybe because we have the ability to do the hard computation to generate such relationships on understanding is dynamic kinds of changes. So do you think that that's a big piece? Once we get the dynamism, are we then done? Now we're actually going to be able to see how the phenotypes genuinely arise? Because it's changes in these networks? Or what's your take on that?
Rosemary Braun 12:02
Yeah, no, I think that collecting dynamical data is crucial. We have been working with, you know, kind of static snapshots for a very long time, or very sparse time series of data, where you can't really see the actual dynamics of gene expression over the course of the day, if we're thinking about circadian rhythms, or the course of development, or the course of a disease progression. And we know that living systems are inherently dynamic, right? So trying to make dynamical inferences from just a static snapshot of, you know, what is the cell doing at this particular point in time when I extracted it from a patient's tumor, for example, right? We're very limited when we're working with that type of data. Do I think we'll be done once we have dynamical data? No.
Art Woods 12:51
Because when are we ever done?
Rosemary Braun 12:52
When are we ever done? Exactly. But the more and more data that we collect the higher and higher dimensionality of that data, we then have to analyze, right? If you think about this, you know, going back to this notion that the whole may be more than the sum of its parts, even if what I'm trying to do is just look at sums of parts, just collections of what particular genes are doing, right? If I'm looking at that now, not just in a static snapshot, but several snapshots over time, or I'm looking not just at gene expression, but also protein, methylation patterns. If I consider all the possible combinations that I might be interested in, I have a combinatorial explosion. There's no way that we will have enough samples to really probe that entire combinatorial space. And so it demands, you know, the development of new mathematical and analytical tools that can extract information from that type of data given how undersampled it is, with respect to how many possible interactions could be taking place.
Art Woods 14:06
So another major focus of your lab, and this is also something that NSF obviously cares a lot about, are the so-called "Rules of Life." And I feel like we've already started to touch on some of those rules, but maybe, just say, how do you perceive what it is the NSF is talking about when they say rules of life? And like, what's a rule that your lab is going after?
Rosemary Braun 14:31
So I think in general, when they're thinking about rules of life, I think what many people think about rules of life, right. What we're trying to get at is what connects what we observe at this microscopic scale to what we observe at the macroscopic scale, right? So if I tell you here is a DNA sequence for a given organism, right, how does that DNA sequence then dictate the shape of that organism, its behavior, its capabilities, its constraints?
Art Woods 15:02
Sort of a gene to phenotype map, would you call that a gene to phenotype map?
Rosemary Braun 15:06
Exactly, yeah. And so I think that's one of the things that NSF is looking for in terms of rules of life. I think another thing that NSF is interested in, in terms of rules of life, is also what governs the what constrains evolution of organisms, right? There are biochemical and biophysical constraints on what a cell is capable of doing. How does that shape the forms that life can take? And how does it constrain the ability of organisms to adapt to changing environments? So I think all of those fall under this notion of "Rules of Life."
Rosemary Braun 15:43
You asked about, you know, what are the rules that my group is interested in, specifically. And so one of the things that we have gotten involved in is a lot of circadian rhythms research. And, I will say that I got into circadian rhythms research somewhat by chance, you know, I met someone who turned out to be a great collaborator, I wasn't doing circadian rhythms work at all, but he had this question, he wanted understand whether we could use the patterns of circadian gene regulation to understand metabolism in fruit flies. And, you know, he started talking to me about it. This sounds really, really pretty interesting, we started working together. And that's how we got drawn into circadian rhythms research. But the reason that I stayed, the thing that fascinates me about it, and the reason that I think it's a really interesting approach to understanding rules of life, is that we have this problem where you need to be able to coordinate biological processes, across cells across tissues, in fairly large organisms, right, like, somehow what is going on across my body, right, a meter and a half of of space needs to be coordinated. And one way for organisms, multicellular organisms, to enact coordination is for every cell to be keeping time. And indeed, that's exactly what happens. So every single cell in your body has a circadian oscillator, it will maintain a 24 hour rhythm, even in the absence of eyes to see the sunrise or a brain to process that information. It's an inherent molecular oscillator. And that oscillator enables cells to orchestrate biological functions at specific times of day that are specific to given tissues, and allows organization to take place across the body without the need for communication, right? If I know that right now, it's 10:20 in Chicago, and you also have a watch, that's enough for us to coordinate when we're going to meet up for lunch. We don't necessarily need to be in contact the whole time, right? And so I view the circadian rhythm as being one way for organisms to enact or to orchestrate biological functions. And that is a Rule of Life.
Art Woods 18:13
So if I can paraphrase you're saying like different cells and different tissues in our bodies, because they all share a common clock don't have to communicate all the time about what time it is. They just, they all know from their clocks, and so they can coordinate their actions. Exactly. Got it.
Marty Martin 18:29
Yeah, well, it's even even more intriguing to me in the sense of melatonin, right? You have the sort of coordination of all the clocks by something that gets into the blood and moves around and ensures that everybody is playing the same game?
Rosemary Braun 18:41
Yeah. Because you need them to be synchronized, right. So they do need to communicate, at some point right, once a day to remain synchronized. But they don't need to be in constant contact about the time.
Marty Martin 18:52
Yeah, you know, I sort of grew up in a lab, I did a postdoc in a lab that was really into circadian rhythm regulation. And it always, you know, it was always in the back of my head, that there reasonably would be other systems in big organisms that, you know, have the same criteria that we're talking about with biological clocks, that there are maybe as a function of natural rhythms, but just in general, this sort of preparedness of tissues for stimuli or stressors or good things, bad things that are going to be experienced in the environment. What are your thoughts about that? I mean, are biological clocks, the only oscillators that exist out there? Or might there be others that we should be looking for?
Rosemary Braun 19:32
We think that there are others too. The evidence is not as strong for things like circannual clocks, that would anticipate seasons, or circalunar clocks that would anticipate phases of the moon, which are important for organisms that live in tidal environments, right, because they would like to be able to anticipate those environmental changes. So there is some evidence that those exist. They're not as well studied as the circadian rhythm, but there is work in that domain as well.
Art Woods 19:59
So I had a question about that just thinking about your network approach to understanding phenotypes like circadian clock-keeping. Do these other cycles, do they use the same kind of fundamental network? Can it get co-opted for a lunar cycle or a tidal cycle or any cycle in the environment? Or are they constructed de-novo for each different cycle?
Rosemary Braun 20:22
So I don't know the answer to that question. I want to be careful here that I'm not sure whether that I don't know or whether it's not known. Because those other, those longer clocks are not as well studied. So I don't know that it's possible that we don't don't know.
Art Woods 20:36
Yeah, fair.
Rosemary Braun 20:37
The circadian rhythm, that mechanism is well conserved in the sense that many organisms ranging from fungi all the way up to mammals have it, the exact molecular underpinnings of that clock do differ from one organism to another. Across mammals, it's basically the same clock. It's quite similar to the clock in the fruit fly. But if you think of cyanobacterium, it has a circadian rhythm, it has a clock. But the molecular mechanisms of that clock are completely different from the molecular mechanisms of the clock in the fly.
Marty Martin 21:13
So I think it's a good time, Rosemary, to maybe talk about Time Machine, because we've been dancing around this really amazing new approach that you've taken. So tell us what that is, and why your group has developed this. It's obviously not the first kind of HG Wells time machine that might be coming into people's heads. It's a slightly different version.
Rosemary Braun 21:33
A slightly different version, yeah. So the name is a portmanteau of time and machine learning. And here was the basic idea. So we know that every cell in your body maintains a 24 hour clock. We also know that circadian rhythms and circadian disruption is crucial to health and disease risk, right? So disruptions in the circadian rhythm have been associated with cardiovascular disease with neurodegeneration with metabolic syndrome. One of the troubles, however, is that even though there's ample epidemiological evidence that links circadian function to human health, we still don't understand how those things are linked mechanistically. And one of the reasons that we don't fully understand the links between circadian rhythms and human health is that it's actually really hard to measure an individual's circadian rhythm.
Rosemary Braun 22:35
So everybody has a natural clock that gives rise to when we feel like getting up in the morning when we have the freedom to set our own schedule. It tells us when we are going to be hungry, it tells us when we're going to feel sleepy at night. And you can measure this and the typical way to do that is by taking serial samples either of blood or of saliva every half hour or every hour, over the course of an entire day and well into the night and probing the melatonin level. At about two hours before you start to feel sleepy, melatonin levels in your bloodstream will start to rise. And the point at which your melatonin level crosses 25% of what will eventually be its peak is considered dim light melatonin onset, it's your clock chiming midnight, it's the cue right? And so we consider DLMO or dim light melatonin onset, as the gold standard marker of circadian phase. So that might be different for me than it is for you. And that would determine how my circadian rhythm differs from your circadian rhythm.
Rosemary Braun 23:48
But like I said, You got to take samples every half hour, every hour over the course of many, many hours so that you catch when it starts to rise, and also so that you catch its peak. Right, because our definition of dim light melatonin onset is 25% of the peak. So you need to keep on going. Usually it doesn't peak until the middle of the night sometime, right, when you would like to be asleep. So this is pretty burdensome for patients. It's expensive because you have many, many samples that you then need to process as a result, nobody does it, unless you're very specifically interested in sleep and circadian rhythms, right, you're not going to be measuring this. So if you're conducting a study on the response to a drug, or you're conducting a study on cardiovascular disease risk, generally you're not going to be including circadian variables in that study just because it's so burdensome. It also means that if you go to your doctor and your doctor is running a battery of tests to to determine your disease risk, or to tell you what time you should be taking your blood pressure medication because the time at which you take it can have a pretty big have a pretty big impact on how effective it is, right? They're not going to include your circadian variables in those assessments, because, once again, it's just too burdensome.
Rosemary Braun 25:08
So this is the problem that we were posed is well. If every cell in your body is running a clock, including the cells in your bloodstream, can we just look at what those cells are doing? Circumvent this whole need for measuring dim light melatonin onset, and get an easier more accessible assessment of circadian function in a human being, right. And so that's what we set out to do. And so the study that we conducted in collaboration with Phyllis Zee and Ravi Aliada here at Northwestern, was to collect blood samples over the course of an entire day, from individuals who had come into the sleep lab, and use that data to train a machine learning model that would be able to use just one sample, from blood, and make an assessment of what time of day that person's body thinks it is. And so now we have a way to assess circadian phase without the need for serial sampling of melatonin.
Art Woods 26:17
So this, this kind of blew my mind when I read your paper on this, that you can get that much information from one blood sample. And I know like machine learning algorithms are kind of black boxes in the sense that it's hard to know always what the computations are that are going on underneath the hood. But what is it about the state of the cellular systems at that one time point that allows you to extrapolate out where you are in this 24-hour cycle?
Rosemary Braun 26:45
Yeah, so the key thing here is that we are using the combination of multiple genes in order to make this assessment, right? So we take a blood sample, but then we're assaying, on the order of about 100 genes, in Time Machine. In our previous work, where we're used to blood samples, we were able to get it down to about 40 different genes. And it turns out that the key factor that enables us to make that inference about what time your blood thinks it is, is the relationship and expression profiles between the genes, right. So over the course of the day, some genes are going to go up an expression other genes are going to go down, and their relative expression to one another, enables us to make inferences about where, over the course of the day, those genes are.
Art Woods 27:36
Yeah, super cool. So do you need information on all of those 100 genes for this to work? Or can you whittle that down to some very small core set, as it were?
Rosemary Braun 27:47
Yeah, so we tried to whittle it down, right, we always wanted a smaller and smaller set. The thing about gene expression measurements is that they tend to be fairly noisy. And so mathematically, right, two genes should be enough to determine the time of day. In actual data, where you have a lot of noise, it turns out that having more genes gives you a more robust predictor. And so that set of roughly a hundred genes is where we ended up great.
Marty Martin 28:17
You mentioned a minute ago, saliva samples, did you take saliva samples? I'm just curious whether your diagnostics for blood do transfer to other cells in the body. If in fact, they are behaving all in the same way that you know, if you only had saliva available, that you'd still be able to do this?
Rosemary Braun 28:34
Yeah, so with saliva, what we were looking at was gene expression, right? So in saliva, you're going to have some cells, not a lot, many of them are going to be microbes. I'm not sure that you could use the same predictor on a salivary sample that we developed for, for blood. But obviously, an assay that would use saliva and could be reduced to a handful of time points, would be much more accessible than blood sample. One of the things that we considered though, when we wrote the Time Machine paper, is that if you wanted to do this in a clinical setting, you could take a hybrid approach, where first you take a blood sample, that very simple single sample thing, run Time Machine to get a rough assessment of the circadian phase for that individual. And then if you really needed to pin it down using the gold standard DLMO measurement, now what you could do because you already have a rough assessment of circadian phase, you could target when you take those serial samples of saliva, do fewer of them, and still get DLMO for that individual.
Art Woods 29:40
Yeah gotcha. So statistically speaking, maybe just tell us so how good is time machine and I see the numbers from your paper. You say there's a median absolute error ranging from 1.65 to 2.7 hours. So like, I don't have a context for understanding whether that's good or not, and how that relates to the differences among people, just in the population in general.
Rosemary Braun 30:07
So to answer the first question, whether it's good or not, in order to be clinically useful, we generally want the error to be, you know, on the order of two hours, ideally less. So Time Machine is kind of right at that boundary. Though a variant where we use two blood draws was consistently under that, right. So if you can afford to take two samples, roughly 12 hours apart, eight hours apart is actually enough. So for example, somebody comes in, in the morning, you do a blood draw, they come in after work, you do another blood draw, then you can get the error rate down under two hours. So with Time Machine, we're kind of right at that boundary of, you know, being useful. But you ask another question, which is, how does that compare to the variability across humans?
Rosemary Braun 30:51
And that is a fascinating question that we actually are still working on. So we've looked at the variability in older versus younger adults. So Time Machine was trained with younger adults, our data there is all from young, healthy, well aligned individuals
Art Woods 31:10
College students
Rosemary Braun 31:12
No, not necessarily college students.
Art Woods 31:14
Oh ok
Rosemary Braun 31:14
But what I believe I want to say that the upper bound for recruitment was 55 years old. And there is an open question about what happens to one's circadian rhythm at the molecular level, as we age. We know that sleep changes, we know that circadian function when we're looking at it, you know, through the lens of, of behavior and sleep and other outputs of the circadian system change with age. We don't yet know what happens at the molecular level, and we've started to gather some data, which will enable us to answer that question. One of the things that we published on a couple years ago, was to look at that that two draw method, the one that was called time signature that used to blood samples, and we asked the following question. So we know that it works really, really well to predict circadian phase in young, healthy adults, how well does it work in healthy older adults? Right, so these are people above the age of 55, the median age, I think, in the older adults was 65, and we had some folks in their 70s in that dataset as well. And what we found was that the predictor, actually even though it was trained with data from younger adults, it worked pretty well for healthy older adults. So you know, I think we still have other work to do, right? We'd like to understand how does shift work impact circadian rhythms, right? What happens in people who have circadian disruption, right, things like insomnia or delayed sleep phase disorder. So those are still open questions that we're looking at.
Marty Martin 32:51
So two more quick questions. And then we want to sort of broaden out to ideas of resilience and such, is Time Machine solely based on clock genes, or what is in your mix of a hundred? I mean, presumably, you went to that down to whatever was most informative. It's not necessarily just clock genes, right?
Rosemary Braun 33:09
Yeah, it's not just clock genes. In fact, of the core clock genes, only one or two. In time signature, so this was the to draw one with a reduced set, only two of them appeared in the final predictor.
Marty Martin 33:26
Wow.
Rosemary Braun 33:26
Here, because we've got a larger set of genes, more of them are included in the predictor. But the majority of the genes that contribute to our ability to detect time of day are actually genes that are related to immune function.
Rosemary Braun 33:40
Aha
Rosemary Braun 33:41
And this is unsurprising because we're looking at blood, right? And what we're looking at is whole blood. So basically all of the white blood cells that are contained in the sample, so it's unsurprising that we're seeing immune markers. We know that the immune system is under circadian control, its activity peaks at the beginning of the day, right when you wake up. So the fact that those genes are showing up is really not surprising.
Marty Martin 34:05
Yeah, interesting. So if you run this forward, and you know, could you envision a day when most clinics and general practitioner's offices are able to use Time Machine? I mean, do you expect that they don't be, you know, timing of surgeries is going to be set on individualized circadian rhythms/ That you know, don't take your medication at any one time of day, take it now because... Is that where you think this is going?
Rosemary Braun 34:28
Yeah, that's where I'd love to see it go. Work from John Hogenesch, and others have really established the fact that drug efficacy changes over the course of the day. We know this for blood pressure medication, which is why in general, you're told to take it in the evening. We know this for chemotherapies. And right now the recommendations for when to deliver those drugs are very generic, right. But imagine if we could measure on an individual basis somebody's circadian rhythm. Now we can give them personalized suggestions for when to take their medications. We could potentially give them lower doses with less side effects and have those drugs be equally effective, because they're taking them at the optimal time. So that's really my dream with methods like Time Machine.
Art Woods 35:18
I think we wanted to broaden out a bit here and talk again about some network properties. And I wanted to ask specifically about how networks achieve resilience. And maybe in your answer, you could address things like how do we even know whether something is resilient? And like, another way to think about this would be if you have a model of a network, can you tell just from looking at that model, whether or not it's resilient? Or do you have to run it and poke it and prod it in order to figure that out?
Rosemary Braun 35:53
Boy this is a big topic.
Art Woods 35:53
Yeah
Rosemary Braun 35:53
Let me see where I want to start.
Art Woods 35:31
So pick and choose which path you want to go on.
Rosemary Braun 36:06
So let me talk a little bit about the way that I think about networks and this question of resilience and how this comes into the analyses that we do in my group. So, you know, we know that genes produce proteins that interact with one another. We have some knowledge about the structure of the gene regulatory interaction networks that govern the expression of those genes, the protein-protein interaction networks, that govern how those proteins interact with one another. So we have these rough roadmaps, right. And you can go to pathway databases like KEGG, or like Reactome, and pull them up. And you'll see this wiring diagram of what's interacting with what.
Rosemary Braun 36:54
The typical way in which we have gone about historically analyzing gene expression or other types of "omic" data in the context of these networks, is first to look at each gene in isolation, identify the genes that are, for example, differentially expressed or differentially methylated are affected in some way. So now you have the genes that differ in your cases versus your controls. And then having identified those genes, you now go to the network, and you ask, "Are they are more affected genes on this network that I would expect by chance?" Or "are they more localized on this network than I would expect by chance?" And this type of approach, it's been useful, right? It tells us something about the graph, or about the expression of genes in the context of the graph. But I think that there's something that's missed there. And this gets to that question of resilience.
Rosemary Braun 37:51
Imagine a situation in which you have one gene that is affected in some sort of deleterious way, right, it should be highly expressed, it's not being expressed at a high enough level. And so if you were to look at that gene, in your cases versus your controls, you might see that it's differentially expressed. A healthy cell might respond to this in a variety of ways, right. Perhaps it's able to then upregulate that gene, recover the appropriate function. But there might be a compensatory change that the cell can make by. let's say, increasing the level of some other genes so that the interaction of the two gene products just becomes more likely, right. You can either have one of them highly expressed, or the other highly expressed, but as long as one of them is highly expressed, you're fine, right?
Rosemary Braun 38:43
Now, if you were to look in this type of data, what you would find is that that compensatory gene also looks differentially expressed. And so if what you are doing now is asking the question, do I find more differentially expressed genes than I would expect by chance? You would say, "Yes, I do." Right, this gene is down, this other gene is up, I have a collection of differentially expressed genes. But those two genes are down and up in a compensatory fashion, which might actually preserve the function of the network as a whole, right? And so I want to be able to distinguish between sets of changes that have a cumulative deleterious effect, versus sets of changes that are evidence of some sort of compensatory process taking place.
Art Woods 39:31
I see that are offsetting in some way.
Rosemary Braun 39:32
Exactly. And so that is where we start to use the structure of the gene regulatory network to probe: does it look like we have statistical evidence of something compensatory versus something that's cumulative damage, right? And that's, that's that notion of resilience that we're interested in.
Marty Martin 39:52
So I mean, resilience, I think it has to include time, doesn't it? I mean, doesn't that word inherently include time. Are you talking about looking at the regulatory architecture, just as a single snapshot in time, back to what we were discussing before. If you're comparing this GRN of one person, say and another GRN, there's no time there. Will you expect to see differences? Could you say differences in resilience is there? Or do you have to talk about dynamics within the network?
Rosemary Braun 40:21
Ideally, we also talk about dynamics, right? But most of our data actually comes from static snapshots or just a handful of measurements, right? You don't have a rich, well-resolved time series. You might have snapshots every few hours, if you're talking about circadian rhythms, or, you know, a handful of times over the course of the year, if what you're looking at is the progression of disease in a population. So oftentimes, we're limited in what we're able to do, in terms of probing the dynamics experimentally.
Rosemary Braun 40:55
The really cool thing is that there's been work in graph theory that shows that the dynamics of the processes on a network is constrained by the structure of the network itself, right? I mean, this is kind of obvious when you think about how the structure of a city's streets dictates the dynamics of traffic, right? If there's a road closure of a major artery at rush hour, that's going to severely impact traffic dynamics, right. And you don't need to actually be watching the traffic dynamics to know that. It's enough to know that a major arterial street has been closed. And so what we are trying to do now is to use those mathematical techniques that relate the structure of networks to their dynamics in order to make predictions or inferences about the dynamics, even when we can't directly observe them.
Marty Martin 41:48
Right, right. Okay. So one more quick follow up. Can you say then, Rosemary, how resilience relates to robustness, for you? Is just robustness, just not something you talk about? Or do they overlap? Or what do you think there?
Rosemary Braun 42:01
Yeah, I think we tend to think of them as slightly different processes related, but different, where robustness is, in some sense, an insensitivity to perturbations, whereas resilience is an adaptation, a change that comes in after the fact to accommodate a change in the environment or an internal change as well, right?
Art Woods 42:25
Yeah, I also had a question about time, but in a different sense, and that is thinking about the robustness of circadian clock keeping. So here's a sort of thought experiment, you know, many things that have circadian rhythms obviously, are small ectotherms, and their body temperatures go up and down. And that feels like it ought to be a just a gigantic perturbation of any sort of timekeeping mechanism, because, you know, what, what do higher temperatures do? They make all of the enzymes go faster, they make the binding between different proteins altered. So how can a clock keep time despite that variation in external temperature?
Rosemary Braun 43:06
Yeah, so this is a phenomenon known as temperature compensation, and in fact, it's a defining feature of circadian rhythms. If you identify a clock, or what you think is a clock, in order to demonstrate that it is truly a clock, you need to be able to demonstrate that its period is constant across a range of temperatures. Otherwise, it's not going to function properly, right, because of what you just brought up.
Rosemary Braun 43:30
Now, the phenomenon of temperature compensation, how that is achieved mechanistically is still not fully understood. We do know that some of the core clock genes have temperature dependent splice variations. And we think that that temperature dependent splicing is actually what enables the clock to adjust, so that it maintains a constant period across a wide variety of temperatures, right? So essentially, it needs to be able to slow down processes that would run faster at higher temperatures. But, you know, the details of how that is achieved are still a little opaque. But we know and this kind of gets to that question of resilience versus robustness. We know that the clock is not temperature insensitive, right? So its temperature compensated in that its period will remain very steadily 24 hours across a wide range of temperatures, but it will respond to a temperature cue, so you can put organisms in constant light environment but have a 24-hour temperature fluctuation, and those organisms will entrain to temperatures.
Rosemary Braun 44:40
We also know from studies in fruit flies that, so fruit flies tend to have a morning and evening peak of activity. Right, i f you've ever sat outside with a glass of wine at dusk, you know this because that's when they all collect on your wine glass. And their morning and evening peaks of activity will change at different temperatures. The evening peak activity is a little bit earlier when the temperature is lower. And we see that reflected in gene expression data from the fruit flies as well. So there's a morning peak of gene expression activity and an evening peak of gene expression activity, and that lower temperatures, those peaks move closer together. So there's both a sensitivity to temperature in terms of controlling, when genes are expressed, but also a robustness to temperature, in controlling that whenever they are expressed, they do so 24 hours later.
Art Woods 45:36
Yeah, neat. I'm not necessarily an adaptationist here, but what you just described about fruit flies, sounds like it almost could be adaptive shifts in sort of the peaks of timing, that are sort of integrating information about both the temperature and the light period, right? So you can imagine, like, it could be good for clocks to be flexible, at least on the short term, to drive, you know, important sets of behaviors over a 24-hour cycle.
Rosemary Braun 46:02
Yeah, and what's really fascinating is that when we've looked at like that morning peak and the evening peak of activity, what we find is that at lower temperatures, those peaks move closer together. So even though the days for these fruit flies in the lab are not getting shorter, right? Morning and evening are still 12 hours apart, as far as the light levels are concerned. At lower temperatures, they're behaving as if morning and evening are closer together, as if the days are shorter. So we think that there might actually be some sort of seasonal adaptation that we're seeing in these fruit flies
Art Woods 46:43
The clock itself, yeah. That's really neat. I think this is a good segue also to talking about gene surrounder. So this is another analytical tool that you've developed. You say "it's a method that integrates expression data and network information in a novel procedure to detect genes that are sources of dysregulation on a network." So unpack that for us.
Rosemary Braun 47:05
Yeah, so you know, going back to thinking about our approach to analyzing this data, where you take the genes, you find the ones that are differentially expressed, you overlay them on the network, you're going to get many, many genes that are differentially expressed. Some of them are going to be the sources of differential expression, others are just going to be downstream of the genes that have changed in their expression level. Right. And they're sort of they're along for the ride, if you will. Our goal with gene surrounder is: could we pinpoint the specific elements in the network that were the epicenters, if you will, of gene disruption, or misregulation, so that those could be targeted? And the reason that we went down this path was that we do a lot of work in these network level analyses of transcriptomic data, which gives us a view at the network scale of what's going on in our samples. But it can be very difficult to then take those inferences and turn them back into an experiment that could be used to validate or to follow things up, right? At the end of the day, our experimental tools are not perturbations of a network, they're perturbations of genes, right? So I go to my experimental colleagues, and they say, okay, great, which genes should I CRISPR? Right.
Art Woods 48:26
You're like, "the whole network. Just do it all"
Rosemary Braun 48:26
Yeah, they told me to get out of their office. I don't know why. So we wanted to be able to start at the network level and then winnow down and find elements of the network that maybe were good candidates for targeting. And so that's what genes surrounder is designed to do. And the way that it does this is that it looks for patterns of differential expression on the network, where you have a single gene that is highly differentially expressed, and then when you look at its nearest neighbors in the network, they're also differentially expressed, but not as strongly. And you go out one step further, and they're also differentially expressed but less strongly still. And so essentially, what we're looking for, is those genes at the center of that pattern.
Marty Martin 49:18
So are those hubs? I mean do you recognize those as hubs? Or are there names for their position in networks?
Rosemary Braun 49:25
Yeah, so usually, when we think about hubs in a network, that's going to be a node in a network that has a connection to many, many other nodes, right. And that's a property of the network itself. The genes that are identified as central genes and genes surrounders might be hubs, but they don't necessarily need to be, right? So they might be a gene that has actually few connections to other genes, or maybe it itself is connected somewhere along the line to a hub in the network, but it is the most strongly differentially expressed, and therefore we think it is the source of the differential expression over the whole graph.
Art Woods 50:05
Gotcha. If I can try to draw the link here to therapeutic uses, which you've kind of already alluded to. It sounds like the power of this is, you know, if you have some complex network of things doing that has some kind of dysregulation or contributing to a phenotype that's undesirable, that instead of worrying about the whole network, you could use this approach to find the sort of key leverage points within that network. And that really simplifies what one might try to do therapeutically. Is that right?
Rosemary Braun 50:22
That's right, yeah.
Art Woods 50:27
Right. And is it seeing uptake? I mean, are people using it in this way?
Rosemary Braun 50:42
That's a good question. I think, you know, a lot of these methods. They're slow to be adopted.
Art Woods 50:49
They start off very theoretical, right? Yeah.
Rosemary Braun 50:51
Yeah so we shall see.
Marty Martin 50:54
So I think, Rosemary, you call these loci regulatory QTLs, correct?
Rosemary Braun 51:01
Yeah. So our regQTL analysis came from a different project than gene surrounder. So our analysis, there was different and happy to talk about the regQTLs, yeah.
Marty Martin 51:13
Yeah please.
Rosemary Braun 51:13
Yeah. So there, what we were interested in understanding is how does the genetic code, what is contained in DNA sequence information, change how a gene might be regulated by other elements in the gene regulatory network. So your audience might be familiar already with microRNAs, these are small non-coding RNA molecules that are thought to regulate the expression of coding RNAs by either binding to them and preventing their translation into protein or tagging them and inducing degradation of the target RNA. And because microRNAs are so short, they're typically six to eight nucleotides long, they bind their targets very nonspecifically. So a given microRNA might have hundreds of RNA molecules they could, in principle, attach to and regulate, right and a given gene might be targeted by multiple microRNAs. And we were interested in understanding what is the relationship between microRNA expression, the expression of genes, and their sequence, right? And could changes in non-coding regions of the genes actually change whether or not a microRNA is able to regulate it, right? So you have a change in the genetic code, the microRNA now can or cannot bind, whereas without that change, it was able to bind or not able to bind, right, so it changes whether the microRNA can target it. And as a result, now you have this regulation by the microRNA, that maybe you don't with a different genetic variant. And so we obtained from TCGA, The Cancer Genome Atlas Project, data for genetic sequence, microRNA expression, and gene expression. And we looked for statistical evidence that there are genetic variants that actually can change how a microRNA regulates a gene. And we found it,and those were what we termed regQTLs or regulatory quantitative trait loci, these are genetic variants that control regulation of genes by microRNAs.
Marty Martin 53:40
Okay, so, I mean, the regulatory QTL, that's a very broad word for a specific effect of the microRNAs. How do you think about the sort of prominence, the importance of these in the context of gene expression, regulation broadly, relative to you know, sequence variation and promoters or CPG sites and methylation influencing expression? I mean, why put an emphasis on this? Is there precedent in the cancer world? Or do you think this is just an underexplored part of gene regulation?
Rosemary Braun 54:13
Yeah, so we were, I mean, we got into this because we're interested in microRNAs and how they work and what is the source of that non-specificity, right? So we know that they're, that the targeting of microRNAs to genes is nonspecific, because they're so small, but evolutionarily, why would that be useful, right, that was our question was you have this nonspecific targeting, why would you want that in the first place and and what consequences does it have for an organism? And so our conjecture was that maybe this non-specificity enables microRNAs to exert redundant control on multiple genes in a pathway, and through that redundancy, allows microRNAs to give more robust control of gene expression at the network level versus at the single gene, right? So if it's not able to regulate one gene, maybe it's able to regulate the rest of the genes, right? And so that was how we got into it, and then we started to say, "Okay, well, if that's our conjecture, do we actually think that there will be instances in which a microRNA is not able to target a given gene?" And so that's why we went down this path of looking for these regulatoryQTLs.
Art Woods 55:35
Yeah that's really neat, because it almost strikes me as if that's the kind of logical inverse of the approach in gene surrounder, right. So there, you're thinking about sort of central things that exert a lot of control. And here you're talking about much more diffuse control over the entire network, potentially at the same time by these regQTLs right?
Rosemary Braun 55:54
Yes.
Art Woods 55:55
Cool yeah, got it. All right. Well, that might be a good place to start wrapping up. It's been a great conversation. But we always ask our guests before we go, if there's anything else you'd like to say, or anything, we didn't cover that you wanted to bring up?
Rosemary Braun 56:08
No, this has been great fun, and we've covered so many topics that I'm at a loss for what we didn't cover. I will say this, you know, because we are a purely computational lab, we really thrive on collaborations with experimentalists. And so, you know, I would say to young listeners out there who might be interested in computational biology, or to those who maybe aren't decided whether they want to go towards experimental biology or the computational route, I think there's so much fun to be had at the interface between computation and experiment. And so I would encourage people to explore that interface. And I think, right now, we're absolutely drowning in data. And we have absolutely extraordinary mathematical and computational tools, from deep learning and very powerful computers, and I think we've only really begun to scratch the surface of what is going to be possible to do with the vast amount of data that we've collected.
Marty Martin 57:12
Excellent. Good times ahead. I like it.
Art Woods 57:15
Fantastic place to leave us. Well, thanks so much, Rosemary.
Marty Martin 57:18
Thank you.
Rosemary Braun 57:19
Thank you.
Marty Martin 57:20
Thanks for listening. If you like what you hear, let us know via X, Facebook, Instagram, or just leave a review wherever you get your podcasts. And if you don't, we'd love to know that too. Write to us at info at big biology.org
Art Woods 57:46
Thanks to Steve Lane who manages the website and Molly Magid for producing the episode.
Marty Martin 57:51
Thank you as well to Dayna De La Cruz for the amazing social media work. Keating Shahmehri produces our awesome cover art.
Art Woods 57:57
Thanks to the College of Public Health at the University of South Florida and the National Science Foundation for support.
Marty Martin 58:03
Music on the episode is from Podington Bear and Tieren Costello.