Scientist Stories: Julie Theriot, Predicting Biology

Now, as I said, we have a packed agenda, so I'd like to get started and introduce our first speaker today. It's my pleasure to introduce Dr. Julia Theriot of University of Washington. Now, Julia attended the college— attended college at MIT, earning a dual BS degree in physics and biology in 1988. She completed her PhD in cell biology at University of California, San Francisco, and then returned to Cambridge as a Whitehead Fellow at the Whitehead Institute for Biomedical Research. In 1997, she joined the faculty of Stanford with appointments in the Department of Biochemistry and the Department of Microbiology and Immunology and is currently an investigator of the Howard Hughes Medical Institute. Her lab moved to the University of Washington in 2018, and she also has an appointment at the Allen Institute for Cell Science. Now, Julia has won numerous awards for her research, including the Packard Foundation Fellowship for Science and Engineering, and the MacArthur Foundation Fellowship. She's also received multiple teaching awards for both MD and PhD students at Stanford, and she is a co-author of the textbook Physical Biology of the Cell. Her talk today is titled Emergence of Cell Shape and Movement from Molecular Scale Dynamics.

Julie. Good morning, everybody. Thanks for showing up this early in the morning. I'm very excited to be taking part in this symposium and to have the chance to tell you about a very new ongoing project spearheaded by Rikki Garner, who's a graduate student in my lab and who I hope is in the audience somewhere. Yes, thank you. So she'll be able to take all of the difficult technical questions. And on the modeling aspect of this project, we've had a lot of great input from our collaborators, Lena Koslover, who's at the University of California San Diego, and Andy Spakowitz at Stanford University. So an overall interest of our group is trying to understand the mechanics underlying the processes of cell motility. For example, in this human neutrophil-like cell— this is an HL60 immortalized cell line that Ricky made with expression of GFP actin— you can see as the cell crawls along, there is intense accumulation of actin right at the leading edge of the cell. And it's generally accepted that it is the polymerization of actin actually impinging against the membrane right at the cell leading edge that actually gives rise to force responsible for forward movement in this system. Now, actin, of course, is a small globular protein that's able to bind and hydrolyze ATP. It's also able to self-assemble to form long helical filaments.. And at the leading edge of a motile cell, such as you just saw in this movie, or which you can see a little bit better here in this structured illumination micrograph, at the leading edge, the actin filaments actually form a dendritic network where the nucleation of new actin filaments grows from the sides of preexisting actin filaments by the action of the Arp2III complex, which forms branches at a characteristic 70- to 72-degree angle. And in this diagram from about 15 years ago, this shows basically based on in vitro biochemical work and a little bit of in vivo immunofluorescence what's generally thought to be the overall molecular consensus for what's happening at the leading edge. We have these branched actin filaments growing. They're pushing up against the membrane. The membrane is acting as a load, pushing back and limiting the polymerization of the actin filaments. And it's sort of the struggle between growth of actin filaments and fluctuations of the membrane that enables the cell to move forwards overall. In order to explore this in a little more detail, Ricky has worked out methods for observing the motility of these kinds of cells at very high spatial and temporal resolution. So in this movie, for example, we're looking just at a phase contrast image of one of these neutrophil-like cells with images taken every 50 milliseconds. And one thing that's very striking here is that as the cell is moving— and this is effectively real time is what you're looking at here— the leading edge actually remains very, very uniform, strikingly uniform, even though what we're talking about here is hundreds of thousands of individual actin filaments that are all growing stochastically. So this is an interesting question as to how all of those many hundreds of thousands of actin filaments are able to coordinate their polymerization So it can all happen at the same speed, even though they're separated over very large distances at the molecular scale. So zooming in on the leading edge here, you can see Ricky's been able to extract the position and the shape of the leading edge, again, at very high resolution. Here we're showing lines just about every 2 seconds, and here, something closer to the actual resolution level of the data. And basically, what we want to be able to do is try to take the fluctuations that we see there and see if we can learn how to understand them in the context of force generation based on this interaction between actin filaments and the membrane. So based on this high-resolution imaging, Rikki can extract not only the position but also changes in speed, which you see sort of flicker in various different regions over the leading edge with sub-second time scales. And those small changes in speed actually can manifest themselves as changes in membrane curvature that persist for longer. So the idea is that the actin is growing a little bit faster in one region of the leading edge. That will cause a region of positive curvature relative to neighbors.. If that polymerization slows down, that will cause a region of negative curvature, and we can track these curvature fluctuations over time. So Rikki basically set off to see if we could take these kinds of very careful observations, these careful measurements done in vitro, and harmonize them with what we understand to be the fundamental physics of the process that's going on here, which is the interaction between these growing actin filaments and the fluctuating membrane. So in order to dissect this mechanistically, we really have to pay the most attention to these two components. The actin filaments in this dendritic network are growing at some angle with respect to the leading edge, and the net amount of actin monomer present in the cell means that there is a favorable energetics for actin monomers to add on to those filament tips. But in order for that actin monomer to sneak onto the filament tip, there has to be space that's opened up, which is thought to come about by fluctuations in the position of the membrane. So we need to understand the physics of the membrane and also the physics of the actin filaments in order to understand how this large-scale coordination can possibly take place. So starting with the actin filaments, Rikki's been able to measure these curvature fluctuations that I talked about and also combine them with fluctuations in the actin density as measured with this GFP-actin cell line. And I hope you can see by looking at these chymographs that there does seem to be some similarity in these two patterns. And in fact, analyzing many cells and pooling all the data together, what she could see is that there's actually this interesting negative correlation such that the parts of the leading edge that have a positive curvature, so those parts that are growing faster, are growing faster than their neighbors actually have a lower actin filament density than the rest of the leading edge. Similarly, the parts that are growing slower than their neighbors, the parts that have a negative curvature, those actually have a higher actin filament density. And this might seem a little bit counterintuitive because you would think sort of from first principles that if you have more actin filaments that are working together to push a local region of the membrane, that they should be able to go faster because they're sharing the work. But this kind of negative correlation actually has been observed in a couple of other systems driven by actin polymerization., including Listeria monocytogenes polymerization and movement of other actin-driven cell types such as fish keratocytes. We'll come back to these examples a little bit later on. Okay, so that's the actin. Now, how are we gonna think about the membrane? Well, our assumption is that because the membrane is under tension, it's got some rest state, and then changes in the position or in the shape of that membrane are going to relax over time. But there can be changes that are induced by local variations in stochastic actin polymerization perhaps in other kinds of fluctuating processes. And in order to be able to dissect these quantitatively, the most simple thing to do is to take whatever fluctuations we see in the membrane and then decompose them using a Fourier transform so that we can look at how the magnitude of these changes evolves over time, and specifically look at how the alterations away from the resting state will relax back to the resting state. And there's been a lot of very fundamental physical modeling done on these kinds of problems where if you think of the membrane as being a series of beads on springs, if the stretching of the springs is what dominates the system, then you expect that the magnitude of these autocorrelations will fall off as the square of the wavelength, and also that the time course of that decay will fall off as the square of the wavelength. But if it's bending that dominates this behavior, then it should go as the fourth power rather than the square. So Rikki initially set up to see if she could reconcile either of these two concepts, either that stretching of the membrane is what dominates the time scale or bending of the membrane is what dominates the time scale. So she took her high-resolution data, decomposed these fluctuations in position into all of their different spatial frequencies by spatial Fourier transform, and then plotted the autocorrelation of those so we can see how these decayed over time. And what she saw here is obviously very different from what we were expecting to see. Remember, these are our two simple models. We expect there to be stretching or bending, but instead what she actually saw was oscillations. So this immediately raises the question, well, where, you know, where are the oscillations coming from? So Rikki, in order to look at this systematically, then went ahead and said, let's figure out better ways, more complete ways of characterizing the nature of these oscillations quantitatively. So she basically took each of these different patterns from each different wave number and then decomposed them as assuming they would have an exponential decay, and then also a sine wave superimposed on that. And doing this decomposition, again, for an average of 67 cells, she was able to measure these scaling patterns for the amplitude and the decay rate of the exponential component, and also the amplitude and the frequency of the oscillatory component, again, averaged over a number of cells. And if you're sitting in the front row, you might be able to see these tiny, tiny little red bars that are inside of the dots. Those are the standard errors. So this is the kind of data that we're trying to match, observing these particular scaling behaviors for amplitudes and time constants, either decay constants or frequencies, as a function of wave number. Okay, so what kind of models can we use to address this quality of quantitative data? Well, remember, we know that there's a bunch of things that are important in this system. We have to pay attention to the membrane stretching and bending. We have to pay attention to actin filament growth. Also to this branching process, and the filaments also can be capped, their polymerization can be terminated. So those are all things that we know about, but it's not at all clear how just the interaction of those particular elements could ever give rise to oscillations. So Rikki began systematically to exploring what other kinds of features we would have to add into our modeling of the system in order to get oscillations to emerge. And in general, the approach that she's taken has been to do 2-dimensional simulations where all of the different elements are treated explicitly. And where the timing, the time steps in these oscillations are scaled to diffusion and to the size of the actin monomers and the thickness of the membrane in such a way that the net speed within the simulation is scaled to the experimental measurements. So we should be able to directly compare quantitative output from the simulations to our actual experimental measurements. She went through a bunch of different possible scenarios to try to find what kinds of additions to the system might give rise to oscillations, and most of them do not. However, she did find two plausible models that do give rise to oscillations. The first one I'll describe is tethering. This was something that was actually suggested in the Listeria system some years ago. And the idea here is basically that you have these branched actin filaments growing up against this fluctuating membrane. Everything is bouncing around, but every so often those actin filaments can get tethered to the membrane, and there's a wide variety of different actin-binding proteins that are known to cross-link actin filaments directly to membranes. That tethering would locally suppress the fluctuations, both of the filament and of the membrane, very transiently., and so the untethered filaments would be expected to polymerize a little bit faster. Now, for these filaments that are held on, that's going to then prolong the proximity of that filament and also its daughter filaments to the membrane, increasing the likelihood of tethering. So we'll have a little bit of positive feedback here, and then at that point, this bunch of tethers you would expect to snap and then move forward quickly, and that could give rise to oscillations. Okay? So Ricky instantiated that in a model, was able to see really nice oscillations emerge from it. And now again, what we want to be able to do is compare quantitative aspects of this model, and specifically these scalings as a function of wavelength, to see how closely this resembles the complex experimental data that she was able to gather. So here again in black, this is the experimental data that I showed you before, and the pink is the predictions from the model. And remember, this model is outrageously simple. This model is only in 2 dimensions, whereas of course the real system operates in 3 dimensions. And furthermore, we're really only paying attention to the actin filaments in the membrane and not to any of the other mechanical elements of the system. But nonetheless, even though this is a really, really oversimplified model, she's still able to get scaling of the simulation for these, again, really subtle and complicated measurements that really looks generally, I would say, pretty much in the ballpark of what we see experimentally. So that's quite encouraging. But I mentioned there were two models that she was able to find that can give oscillations. The second one is also something that has been suggested by previous work in the literature, specifically work by Michael Sick's group on actin polymerization at the leading edge of a different motile cell type, the keratocyte.. And here the issue is that if you have an actin filament that's growing really perpendicular to the leading edge and actin monomers are being added at some rate, that is actually going to be able to grow faster in this perpendicular direction than a filament that's at an angle to the leading edge. The filament at an angle to the leading edge basically is penalized by cosine of theta because it's not growing perfectly straight. But the downside of growing straight is that if branching only occurs right next to the membrane, which again is experimentally supported by the fact that there are a lot of known factors that activate the branching complex, the ArpG3 complex, right here attached to the membrane. Then the filament that's growing straight, its branched daughters will tend to go off at this characteristic 72-degree angle and then grow away from the leading edge, so they'll get capped and they won't be able to contribute to force generation. In contrast, for an actin filament that's growing at an angle, even though it's growing a little bit slower, its daughters are gonna have about a 50/50 likelihood of actually being able to contribute to pushing on the membrane. So it's pretty clear why you would get differences in speed and actin density with these two scenarios. So in this case, you'll get faster growth but lower actin density. Here you'll get slower growth but higher actin density. Remember, that's consistent with the data that we saw. It's not totally obvious why this would give rise to oscillations, but nevertheless, in a realistic simulation, Rieke was able to find that just making that set of assumptions, you can have variations in the orientation of the actin filaments, is in fact sufficient to give you oscillations. Again, comparing this to the experimental data, the amplitudes are not quite right, the scaling is not quite right, but again, given the oversimplified nature of this, we at least feel like it's encouraging that we're more or less in the right ballpark. So now we have these complex observations, you know, very subtle, very nonintuitive, and then we have a couple of reasonable plausible simple models that at least get us in the right ballpark in terms of how this behavior evolves over time. So how can we distinguish between these two models? Can we determine if one of them is more plausible than another? The way that Ricky has started to go about doing this is to actually perform experimental perturbations where we're disturbing certain aspects of the actin membrane system in the cells and can then go back and rerun the simulations by changing quantitative parameters in ways that we would expect those particular drug perturbations would end up affecting filament growth. So the example I'm gonna show you today is latrunculin B. This is a small molecule, a drug that's known to bind and sequester actin monomers. And this has two effects on the growth of the actin network at the leading edge of the cell. First, it effectively decreases the available actin monomer concentration, so it slows the net growth rate of individual filaments just because there are fewer actin monomers around to add. And it also decreases filament density, again, for the same reason. So with the drug treatments, the drug-treated cells are shown here in green. And remember, the standard errors are these tiny little bars that are inside the dots. So these changes may look small, but they are in fact clearly statistically significant and biologically meaningful. And if we look at the first component of these fits, which is the amplitude of the exponential decay, you can see the effect of latrunculin is basically to increase the amplitude, but then to decrease the timescale.. In other words, you get larger fluctuations that fall off more quickly. And this is, again, sort of intuitive for what you might expect if you have fewer actin filaments that are able to cooperate with each other. On the oscillatory component, again, the addition of the drug increases the amplitude but decreases the frequency now of the oscillations. So now Rikki was able to go back to her simulations, perform them under alterations in quantitative parameters that should imitate this drug treatment, and then be able to evaluate whether or not our simulations can at least match the correct directional change of these various different parameters for this complex system. Whoops. So this is model number 1, which is tethering. Remember, the experimental result for these 4 different parameters is that addition of lechonclinase is an increase, a decrease, an increase, and a decrease. And here in the simulations, we're looking at effectively low, medium, and high concentrations of the drug. Here changing the density of actin filaments for this particular model. And what you see is that for this tethering model, going through this exercise, the addition of drug, in other words, the decrease in the number of the actin filaments, actually causes increases, systematic increases in all four of these parameters. So the sort of scaling is roughly all right, but the direction of change is wrong, at least with respect to these time scales. Now, in contrast to that, this orientation model actually matches surprisingly well. It increases the amplitude, decreases the decay rate, increases the amplitude also of the oscillatory component, but decreases the frequency of the oscillatory component. So this was really, again, very surprising and came out of this careful comparison of really oversimplified models that Ricky designed in order to address what we knew to be the most important components of the system, together with very, very careful quantitative measurement done under conditions where we could really be confident if the changes we saw were statistically significant or not. So to summarize, this has been entirely the work of Ricky Garner, who is sitting here, sitting right over on the edge there, with some input advice on the modeling side, but all of the experimental work was done entirely by her. And what we concluded from this essentially is something that's no surprise to anybody in this audience, that careful quantitative measurements of complex experimental —in this case, leading-edge dynamics in the moving neutrophils—for the neutrophils revealed unanticipated behaviors. And here, specifically, we saw this oscillatory behavior, this regular oscillatory behavior, which is not obvious from any of the movies you guys have ever seen with cell motility because it happens over—only over very short time scales and over very small distances. But it is something that we're able to measure. And the fact that you can see oscillations at the same time that this leading edge is moving forward so smoothly suggests there's probably some kind of force feedback in the system to keep the leading edge essentially smooth, even though it requires coordination of the stochastic growth of hundreds of thousands of actin filaments simultaneously. These stochastic simulations that Rikki has worked out have been— has been used by her to find plausible simplistic mechanisms that can recapitulate these very unexpected features of the data. And in particular, right now, you know, at this moment in time, we think the most likely mechanism comes down to this observation of variant variable filament orientation in the growing actin network that seems best to recapitulate complex and unanticipated aspects of this data. So I will end there. Thank you very much for your attention. I'm, of course, happy to answer easy questions, but then Ricky is here to answer hard questions.

Thank you. Thank you. Yes. So we are going to have microphones that can come to you when you have questions, but first, can I ask people to take a moment, maybe move in a little bit? We have people standing there looking for seats. So please, if you can take that time, and then people in the back, please feel free to come down and grab seats. So we have the microphones ready in the meantime if anyone has a question.

Why only 20 frames a second? Oh, why only 20 frames a second? So we're basically limited by the speed of our camera response. We've been capturing the shape of the entire cell. It's possible that we might be able to go faster if we actually zoom in just on a little part of the leading edge, and that's definitely something we've talked about. About but just haven't moved in that direction yet. For some of the movies, not all of them, but for some of them, we've also been trying to capture the data using a beam splitter right in front of the camera so we can see the fluorescence of the GFP actin absolutely simultaneously with the position of the leading edge by phase. And in this particular line, I didn't go into the details here, but that HL60 line, it's expressing only GFP-tagged beta actin. The genes in the genome actually have been replaced with tagged versions of those using technology that was developed by the stem cell editing group here at the Allen Institute. So every molecule of beta-actin in those cells is GFP-tagged, so we know for sure that the intensity— fluorescence intensity there is a direct readout of actin density. So the only way you can image it is with the fluorescence? Well, that's the only way we can image the actin density. With current technology? With current technology, yeah. For the phase contrast, we might in principle be able to go a little bit faster, you know, if we play around a little bit with things like the size of the field of view so we can get past the limitations of the camera. Thanks.

Only 20 frames per second. In back, blue shirt.

Microphone's coming to you right now. In your fitting, you used a non-decaying oscillatory component in your autocorrelation functions. What is the physical reason to expect like a perfect memory on that? Yeah, so we've been back and forth a little bit about whether we should have there be— should it be a sum between an exponential decay and a non-decaying sinusoidal component, or should we multiply the two? Or you could imagine a couple of different scenarios that would give you sort of generally similar behavior. The argument in favor of taking the exponential decay and then superimposing on that the addition of an oscillatory component basically comes down to the fact that we don't see negative things. It never goes below zero. We never see the leading edge actually regress. This is always just changes in the speed the speed of something that's constantly moving forward. And so that puts a limitation on the multiplicative model that just becomes not an issue, a non-problem in the additive model. Ricky has been playing around with other forms of that fitting. So far, we haven't seen anything that obviously works better than this sort of simple model that we have, but it's not something we can really fully justify.

Ricky, would you like to say anything further about that? Okay, thanks. We have time for one more question in back.

The microphone's coming to you. What's happening at the other end of the actin molecule? Is it being degraded at the same rate? Yeah, so the actin, when it assembles, is primarily in the ATP-bound form, mostly because there's chaperones in the cytoplasm that catalyze nucleotide exchange on actin monomers. And the ratio of ATP to ADP in the cells is about 90% to 95% ATP. So when it comes on, it's ATP actin, and then once it incorporates in the polymer, then the catalytic rate speeds up. So as the polymer ages, you get more and more ADP actin embedded within that polymer. The ADP version of the polymer is less stable in general, and then there's also disassembly factors like cofillin that specifically recognize the ADP form of the filament and take it apart. So there's sort of a treadmilling that goes on, which is biased towards growth at the front and disassembly at the back. That's coupled—indirectly coupled to the nucleotide hydrolysis within the filaments. You may have noticed, though, in the movie and then also the fixed-cell image that I saw that the lamellipodium is really uniform in terms of its density from front to back. And if you measure very carefully, you can see a sort of exponential decrease in terms of the active filament density, which is what you would expect from the simple model I just described. But then when you get right to the back of the lamellipodium, it just falls off a cliff, and you get a very sudden drop in actin filament density. That as far as I know has never been properly explained. We have some ideas kicking around about how there might be interactions with organelles within the cell and specifically with the endoplasmic reticulum that helps sort of tidy up that back end of the lamellipodium, but those are all things that are under active exploration. Wonderful. Thank you very much. Thank you so much.