Got it. Yeah. Thank you so much, Hannah, for the introduction and also having me It's my first time in Atlanta and so far. So again, thank you all for the hospitality and coming to the talk. On Zoom can hear me, okay, I can, Can everyone knows. Alright. So one of the most fascinating things about the brain to me is its ability to interact with the environment in a flexible way. You can think of the brain as an input, output machine, interprets the environment through the sensory system and acts on the environment through the motor system. But this relationship and interaction with the environment is highly dynamic and flexible. And how does the brain, how, how, how did the brain circuits subserve this flexible interaction with the environment? That's really a core interests or that's driving the research in my lab over the last 12 years now, how do you CSD? With this point of view, we work on a number of different topics. I'm here in just listing the three main thrusts of the lab. And the topics that my lab are pursues, motor learning, decision-making and olfactory learning, olfactory perceptual learning. I'm happy to talk about any of them, but for your talk, I have to pick one. And I wanted to share with you some of our latest developments. Looking at brain mechanisms of motor learning. Yeah, please stop me and ask questions anytime I'd rather much rather have this be interactive and understandable than just going through all that, all the slides. So, yeah, please stop me. Alright. What is that? What is motor learning that we study? Well, it's different. There are different types of motor learning. Different people define it differently. A particular type that we are interested in is the learning of motor skills. And motor skills are learned reproducible movements that reliably achieve behavioral goals. And reliability, to me is really the key to what's being learned. So just imagine you're learning to play tennis. And when you are beginning, when you're a beginner, you might be able to hit a really nice job once in a while. But it takes sometimes years of practice to be able to do that reliably across trials. Alright, so something in the brain is changing to allow you to do the same thing, reliably reproduce the same movements. So how does the brain do that? That's, that's, that's the key question that we've been after. Here's one other slide. I used the highlight this topic. So here's two videos of my first daughter. On the left is the moment that she walked for the first time in her life. Then on the right is one year later. And these are all actually are five. This is five years later. So now she's much more capable of performing the more complex locomotor locomotive movements. Two things I wanna do on this slide. One is to really showcase how beautiful my daughter is. But the other thing that's perhaps more relevant to my talk here is that even a simple or a common behaviors like locomotion or learned and practiced through trial and error, so that your movement becomes more reproducible and reliable across time. So that's, that's really something that's really what we want to understand in this line of research. So when we started this investigation, the number of years ago, we started by trying to adopt this question in lab mice. And here, here's one of the tasks that we use to study this. So here the mouse is hit fixed and given access to a liver that they can use to grab onto and move down. Right here. This is the first day that the mouse is performing this task. The gray trace is showing the movement kinematics on individual trials. So those are, the liver is moving over time, overlaid across trials and you can see that there's really no discernible pattern that's reproducible across trials. But through daily training over two weeks, this movement becomes more and more reproducible and reliable so that this reproducible movement pattern emerges as a function of daily training. So hopefully you can see the analog to what I talked about earlier. Let's say tennis swing. Initially your swings are variable, but you develop a more consistent movement pattern, right? So we think we are tapping onto an important key aspect of motor learning that we're interested in. I'm not gonna go into all the things that we did in the past, but one of the things that we showed, I think conclusively, that this initial phase of learning. Requires the primary motor cortex and its activity. And so a lot of our study focuses on changes in the primary motor cortex or M1, during this initial phase of learning. A quick side note, please report. Yeah, so that is an interesting question. This is definitely a reinforcement learning in the sense that they're making this movement to receive the reward. But the fact that they are using different types of movements early on, that all leads to a reward. But eventually by repeated practice, they just pick one of those movements and stick to it. That's the part of motor learning that we're interested in. And yeah, great question. Thank you. And so quick side note is that we've shown also that if we train mice with this task for additional month or two, then this movement eventually becomes independent of M1 in mice can do this task without m1 when it's highly overtrained. But we are interested in the period, early period during which the behavior critically depends on them one. So one of the first experiments we did was to apply a two-photon calcium imaging. This is a technique that we often use to record the activity of a population of neurons. And one of the key advantages of this technique is that we can use it to record the activity of a population of neurons one day, put the mass back in their home cage, take them up the next day and very reliably identify and record from the exact same population of neurons across, across time or for a number of days to weeks. And so we use this technique to study how identified population of neurons, their activity might change during learning of this task. This is something we did a number of years ago and I won't go into too much detail, but just to show you the key findings here. So here, this is the activity pattern of a population of neurons during the beginner phase of learning. Here, different colors correspond to the activity timing of different neurons. And each dot represents a trial at which the neuron was active. And hopefully you can see that at this phase of learning, activity timing is highly variable from trial to trial. But with daily training over two weeks, each neuron begins to pick a favorite time to be active. They are reliably active across trials at the same time relative to the movement onset. So that the whole population forms a reproducible sequence of activity. Alright? And as a result, that the activity, population activity becomes similar across trials, which is quantified as the population activity correlation here that goes up as a function of training. And one of the key aspects of this is shown here. So as a result, learned movement, this expert movement, and we'll come back to this later. But this learned movement is paired with a reproducible sequence of activity. But as I said earlier in the tennis swing, even the beginner can perform the, a nice, generate the nice shot once in a while and I swing once in a while. And we see the same thing in mice as well. Even in the early phase of learning. There are a small number of trials where the most generates this learned activity pattern of learned movement pattern, excuse me. But when we look at those trials, the activity pattern is very different. So what this suggests to me is that there is a target movement that is being learned. There are different population activity in M1 that can eventually generate the movement. During learning, they are trying, trying out different activity patterns until one particular activity pattern is somehow chosen. And something in the brain changes to be able to reproduce that activity pattern so that it is reliably paired with the movement that is being learned. So there's a degeneracy to the activity pattern that is paired with the target movement. And what is happening during learning is some changes that allow the brain to reproduce one of those activity patterns so that it is reproducibly paired with the learned movement. Alright, so that's the quick summary or background for the two stories that I wanted to share with you today. To investigations at very different levels. So first, we're interested in the mechanism of the activity pattern changes and we look at the synaptic level. Cancel specificity of synaptic changes in M1 during motor learning. And in the second part of my talk, I will share with you the, the ongoing investigation looking at the role of this activity pattern. I'll start with the synaptic. Synaptic level changes. And just to reiterate, this is really trying to understand how the population of neurons go from here to here. So very variable activation to reliable activation. Okay? And we think synaptic changes are key to this observed change. We started this by applying high-resolution imaging. This is a structural imaging of z stack of a sparsely labeled m1 neurons. We're looking at the layer, one superficial layer, looking at the dendrites and why you see these little protrusions out of dendrites. Those are dendritic spines or post-synaptic sites of these synapses that these neurons are receiving. And we showed that during the learning of this task, we can identify these synapses are formed during learning, so they didn't exist at the beginning of learning. They formed during learning, after a few days. And also there are some synapses or spines that existed at the beginning of learning, but eventually get eliminated during learning. So there is a turnover of synopsis. Going into these neurons. Showing this in a schematic, we find the distal apical dendrites on the superficial layer. Layer one generates new synapses during learning and eliminate some of the old synapses. So there's a compartment specific turnover of synapses. Long do these M1 neurons. And so this was again a number of years ago. And it's nice in the sense that all gets so new synapses form as a function of learning. But it doesn't really tell us how these synapses contribute to the changes. And yeah, good question. Is there a spontaneous changes? You see some of this by learning enhance these dots. And this is especially during the first five days of the two-week long training. This process is happening very fast. Yeah. Alright. So back then, we didn't know what these new synapses dead, right? We just knew that that happened, but we didn't really know what these new synapses encoded in terms of inflammation or activity. And a new approach became available in the last few years. Mainly driven by the improvements of glutamate sensors called sniffer, originally generated by Roger Chen at UCSD, improved by Laura and Lueger Janelia, who is now also at UCSD. Now, we can express this sensor on the dendrites and report the presynaptic release of glutamate so we can image them at individuals Stein level when those spines are receiving the presynaptic release of glutamate. So basically when the presynaptic neuron is active, they release glutamate and we can image the postsynaptically. And you see that the different synapses are active at different times. Those though they are, they belong to the same dendrite. What did I do with my water? So taking, taking use of this new sensor, we express this in one layer, 23 neurons and let them learn this task and image the same dendrites across, cross the learning period. And identify the new synapses formed during learning and looked at what kind of inflammation they encoded, what basically when are they receiving inputs. Here is 12 examples of new spines that formed. Why are the mice were learning this task? And the activity of each of those spines is aligned to the movement onset shown in the dotted line here. And we were excited to find that these are many of these new synapses we're active, time-locked to when the mouse is performing this task, when the mouse is generating this learned movement. Because the population of those, all of those new synapses, many of the synapse, synapses are active, wrap that right at the movement onset, or some of them before the movement onset, consistent with the idea that these synapses become active that eventually lead to the generation of the learned of movement. And out of, at a first glance, the activity of these synapses look a lot like some of the pre-existing synapses. We call the movement related spines. So these are spines that show increased activity time-locked to the learner movement. And these are 12 example preexisting movement related spines. You can see qualitatively that new spine activity is very similar to the activity of movement related synopsis or spines. So this is good. This is, to the best of our knowledge, is the first time show that the learning induced synaptogenesis leads to synopsis that encoded learning-related inflammation. So that might have been a prediction but had never been shown. And it's nice to, nice to confirm that. I think I saw a hand Yeah. Spatial crusher to the spine. For example. Together. Yeah, that's really the key topic of the rest of this part of the talk. So yeah, Great question there. Hold that thought. Yeah. Yeah. I think we can we can go through the model and then if you have additional questions, we can come back to that. Thank you. Any other questions? Alright. So okay, so new synapses in code and learning-related inflammation, that's good, maybe not so surprising. What was very exciting to us as sort of schematized here, illustrated here. So here I am overlaying the activity of two spines, alright, the blue in new spines, blue is a new spine and red is a nearby unrelated spine. So they are activity, you know, different. But quite often they coincide, they're co-active. There are two different synapses next to each other that receive inputs at the same time. This is more likely when those two synapses are close together. New spines show correlated activity with nearby movement, with movement related spines, especially when they are close together within five or ten microns. So this is the very smallest scale coordination of activity that is happening on the dendritic branch. And for those of you that don't think about dendritic integration all the time. This is an exciting finding because previous work, mostly in slides, have shown that the location of inputs matters. So when multiple synapses near each other or co-active, they can have a super linear cooperation to drive the postsynaptic neuron strongly. So even the same amount of input, shown in three arrows here. What if they are distinct from each other? They only add linearly and they only have a small effect on the postsynaptic neuron. But if they are clustered together like this, they can have a super linear effect on postsynaptic neurons and drive the postsynaptic neuron strongly and robustly. Previous people have postulated that the perhaps learning takes advantage of a mechanism like this. So if Lauren, Lauren, the inflammation is clustered together on the dendritic branch that can lead to a reliable activation of that neuron when those synapses are active. And that might lead to a reliable execution of the learned movements, a learned behavior. We think we have evidence for that. So when these two synapses are co-active, they can have a strong impact on the postsynaptic neurons activity. Now we became really interested in studying what happens when these synapses are co-active. We find almost all of these core activity events happen during movement, during or right before movements. So we only see these synapses being co-active when the mouse is making the learner movement. And what kind of movement are they making? Here is an example, five movement examples coinciding with the core activity of a new synapse and the movement related snaps next. And these are five different movements that don't coincide with the learn the core activity events. One is more reproducible than the other, and this is a highly, highly reliable results. So when the new spine is co-active with nearby synapses, the resulting movement is much more similar to the learner movement pattern. Just to remind you, this is a learning of a task in which mouse is learning to generate a reproducible movement. And when the new synapses or co-active with nearby movement related synapses. There are movements tend to be. Tend to be the reliable, reproducible learned movement. So consistent with the idea those coincident inputs to these new synapse unrelated clusters are driving this learned movements. Alright, so these two snaps is near each other or co-active. How does that happen? How are they co-active? Here? We can think of two simple scenarios. One is that they receive input from the same axon, then you'd expect that they, their activity must be often co-active. Whenever the presynaptic neuron is firing an action potential, there are downstream spines would be co-active as well. Considering potential possible relief failures, etc, which would reduce reproducibility. Perhaps some more exciting finding would be that these new synapses receive inputs from a different axon that is different from the nearby movement related synapse. But somehow these different axons, different neurons, different presynaptic neurons show correlated activity so that the postsynaptic spines would show correlated activity. This might be seen as evidence that the synaptogenesis might bind new information streams onto this dendritic representation, different axons, the new axon nevertheless correlated with the pre-existing connection. This is conceptually simple distinction, very hard to experimentally test because when we are imaging, this is a postsynaptic neurons. We're using a sparse labeling. We cannot see anything beyond what is labeled. And so the only way we were able to think of, to try to get at this question was to perform, What's the performance? In vivo imaging longitudinally, identify these, snap the clusters, and then take out the tissue and perform 3D EM to visualize the tissue surrounding the, the the, the image dendrite. The other question. I think Kathy, we have a question from Zoom, which is from Dieter Yeager. He asked, what are the pre-synaptic sources of co-active axons? Thalamus cortex are out there. What is the source like? Where where in the brain are they coming from? Yes. Okay. Yeah. Great question. We don't know. We're trying to look at that. We have a hypothesis which is some of these likely come from the pharmacy. So we think the phonics emphasises the cortex, sort of kick-start the population dynamics and m1. And at least some of these new synapses that form might be receiving inputs from the thalamus, but we have not shown that. Yeah, Great question. Anything else? Okay. So we did that. We did perform the longitudinal in vivo imaging and took out the exact tissue that we were imaging and surrendered the tissue to this workflow of correlated electron microscopy. Partnering with a local colleague, Mark L. L is mum's group. I'm not gonna go into the technical details, but well, we can do is the following. So we perform this in vivo imaging. We image this dendrite across learning and then took it out and perform a 3D EM. We can manually identify the dendrite that we were imaging and reconstruct this. And it's very convincingly the same dendrite and we can even identify the same synapse across preparations. And now we are able to zoom in. Say, alright, this is a new synapse that formed right next to this preexisting movement related synapse. And we know that these two synapses show correlated activity. And when we reconstruct the presynaptic axons, even though these spines are only a micron apart from each other. We can show that the connected with two different axons that run basically perpendicular to each other. This is almost always the case. It's not always, but almost always these new synapses. So functional clustering with other spines that receive different axonal inputs. We think this is evidenced for evidence that these new synapses are binding new information streams onto, onto the dendritic tree representation. With this and other lines of evidence that we don't have time to show you today, we propose the following model. So we start from a dendritic representation when we are trying to learn something new. And some of those synapses happen to be important for the learned behavior. And then those synapses are used and reinforce to be strengthened, they put, they undergo synaptic potentiation. And the synaptic plasticity is signal diffuses biochemically in the local neighborhood of the dendrite that induces filopodia protrusions, basically dendrites start searching for potential synaptic partners in the neighborhood. This process might be somewhat random, but if some of them happened to connect with axons that show activity that correlates with nearby synapses, then that connection preference really gets maintained or stabilized so that they become stable new synapse. This is a bio, biologically feasible mechanism. So all of these, each of these aspects can be related to specific molecules, etc, that have been studied previously. What is satisfying about this is that this is a simple, biologically feasible mechanism, but that achieves what we want to achieve during learning. So the result of this is that these synapses form of functional cluster that receive inputs that correlate receive correlated inputs. And as I said earlier, when nearby synapses are becoming co-active, the reliably drives the postsynaptic neuron. And we think that might underlie this reliable activation of neurons that we see as a function of learning. So when the, when the mouse is generating this learned movements, coincident inputs arrive to the functional clusters that have been expanding, expanded by these new spines. And that reliably drives the activity of these M1 neurons, leading to the reliable reproducible population activity pattern. Yeah. So that's that's, that's the model that we propose. You mentioned. Do you think there's a biochemical signal from me, strengthened synapse to kick off the formation. What do you think that is and have you tested it? We have not tested that directly, but actually, so that part of the work comes from this first author, Nathan Hendrick. He's a post-doc Laura or excuse me. So he's supposed to be my lab. His PhD work with real, Hey, SDA showed a number of rho, rho GTPases that's that can diffuse within five to ten microns and changed the plasticity threshold of nearby synopsis. So we think something like that is doing what we see here. Yeah. Good question. Any other questions before I move on to the second and hopefully a shorter part of the talk. All right. She'll have a question. Yeah, go ahead. So this is the case for the nearby synapses, right, in the same, same dendritic compartments. But I was wondering if you've ever like actually consider all looked into I'm inputs from different compartments, like different compartmental dendrites in the water. That time delays can actually initiate a sequence of movements. Sequencing with types of movements. Yeah, that's interesting. I'm trying to understand. So, okay. So the sequence is across sales, right? So each cell is reliably active and follows. Some cells follow others. And so we tend to think that those later, later, later neurons are driven by earlier neurons. None necessarily the intracellular computation within the single neuron. But, but it's possible that maybe those later cells are driven with less coincident inputs. And so maybe their activation is delayed. And actually that comes with an interesting observation, right? So these later cells are less reliable, less precise in their activity pattern. So your model could potentially explain that. So I like that, I like that. I haven't thought of it that way, but I think that's possible. We have another question from Zoom. This is from Simon spawn Berg. Yes. If the animal is learning multiple movements, what do you expect? Multiple types of correlated spines in the same neurons or different neurons underlying each of the movements. Both outcomes seem challenging because the first could lead to the growth and elimination of synapses could conflict. In the second case, we would seem to need more. We would seem to need a very large library of neurons. Yeah, great question. I think it's both. At the population level, the different movements activate overlapping but different sets of neurons. So in that sense, the plasticity is likely partially overlapping in terms of which cells undergo through this kind of plasticity. But even within individual neurons, there are mechanisms to compartmentalize this kind of processes. So, for example, different dendritic branches can compartmentalize this. And indeed, when B, again, I had a really nice paper a number of years ago that suggested that when a mouse learns to walk forward versus backward, the synaptogenesis that relates to Florida versus backward are often found in different dendritic branches. So even in the same cells, there are ways to compartmentalize this kind of processes. Yeah. Great question. Alright. So I'll go into the second partner, switch gears and talk about this second projects. And I should say this is a very new project ongoing. First time I'm talking about it in public, so it's gonna be rough and preliminary. And I'm doing this looking for feedback. So yeah, I'm excited to hear what you think. The big question is. Okay, so we, we've shown that this population activity changes correlate with learning. So it happens during learning. And part of the activity happens before the move man. So it's definitely not just a result of the movement, but is it really driving the movement? So that's a question that we haven't been able to address until recently. Again, these things tend to be driven by technological development and technique that became available to us in the recent years is optogenetics stimulation with single-cell resolution. So here we perform calcium, I mentioned, of a population of neurons. And then we can say, alright, I think these red neurons are interesting. So we'd like to activate them and use the spatial light modulator, selectively deliver light, a two-photon light activation light to those cells and activate them very reliably. So these are red neurons that are activated here. And notice the other, other cells are not activated at that time. And we can move the light to now activate these yellow neurons. And then these neurons are active. These are the yellow neurons and other neurons, including the red neurons that I showed at the beginning, are not activated. And we can move to the third stimulation group here, the green, green groups of neurons, etc. So here we have the ability to identify neurons based on their activity because we're doing calcium imaging and then selectively activate the neurons, a population of neurons that we think, interesting man, here. And we can do this in 3D. So we can image different planes and move across plains quickly and activate a subset of neurons in each plane, semi simultaneously. So this is the approach that we use to address this question. Is the activity in M1 able to drive movements? Alright? And we use a very similar task. So here the mouse is held fixed again. It has the liver. Here's a tone. So the task is to press this liver in response to your tone. And here's the mouse that's performing the task. So this left hand is pressing the lever for and receives the water. So it's like you can also see the licking. And we're going to do is in a subset of trials, we're not going to play the queue. We're not going to play the tone to tell them to move, but instead we activate some neurons in M1 using optogenetics. See what happens to the behavior. Alright? So we trained mice with this task and image the activity of a population of neurons and M1 after two weeks of training and reproducing our earlier results, we find reproducible activity patterns of these M1 cells. And some cells start right before movement onset consistent with their role in driving the movement. So we identify those cells. Then in a small subset of trials, instead of delivering the sound cue, we activate those neurons. And here's the result. So here again, I'll show you the same movie on the left. This is a voluntary push. The mouse is pressed into the push, pressing the liver. In response to the sound cue, we're not doing anything externally, or we're not doing anything to the brain. Here's another trial where we're not delivering the queue, but we're activating M1 neurons. And the first approximation, I cannot tell the difference. Alright, they're there. They're pressing the lever exactly how they would in the normal trial. Except that there's no reason for them to press the lever. They are not here in the queue. And also we don't want to reinforce them to press the, press the lever in response to the optimal stimulation so we don't reward them in these trials. So this is a completely on reinforced behavior that is driven by the optogenetics stimulation. And, and, and, yeah, this is only about 20 cells that are being activated by our stimulation or directly by our stimulation. Dieter asks whether these are still L23 neurons in common is that they wouldn't project to the spinal cord. That's right. Yeah, we are doing this all in layer 23, so they are not directly projecting down to the subcortical structures. That's right. Yeah. The recorded neurons you showed on the last slide, the neurons that fire the earliest look pretty close to 0. And you said they were consistent with maybe a premotor ish delay, would you expect that to be closer to like two to 300 milliseconds before the onset of the movement. It looked like they are very close to 0 in the pseudo-colored plot that you showed. Yeah, that's right. So in these trained animals, the reaction time can be like a hundred, two hundred and fifty milliseconds. And that's when these neurons start being active. I see. Okay. And in the case of the optimum induced push, what's the latency that you see between upto and the induction of the movement. Yeah, good question. I think it's about the same one hundred, two hundred milliseconds. Similar to the reaction time as we see in the voluntary, interesting. So do you think if you move to auditory cortex, for example, in stimulated neurons that were active during the tone, you would get a longer latency. Use the same carmine preparation there. Yeah, good question. Yeah. So so at different levels of hierarchy, can we see the hierarchical processing on that would be the prediction. But I don't know if the delay is long enough to be able to result, but yeah, that would be definitely a prediction. I don't necessarily think this is auditory cortex dependent, but that's a separate grade. Pretty cool. Have you stimulated neurons that you don't think are involved in the task? Now, I'll touch on that later. Yeah, we are just starting to do it. We're starting to analyze the results. Not much we can say at this point. Yeah. But, but what I can say, it doesn't have to be the right cell, right cells to generate some movements. But activating the right cells make better movements. And I'll come back to the point. Yeah. Okay, So this was really surprising to me. I didn't think this should work. I think the brands should be like you just activate 20 cells and you generate the alarmed movement. I thought it was a more distributed code across many cells, perhaps across different brain areas, right? But this is really showing that activation, direct activation of neurons is sufficient to induce this, you know, maybe simple but not more complicated than just a twitch kind of movement, right? So how is this happening? We started looking at the movements more closely. Here is analysis of voluntary bush, voluntary trials. So we're looking at different trials across trials, right? So, and we're basically analyzing the quality of the movement. How similar is the movement that's, that's induced in individual trials relative to what we call the learned movement. We identify consistent movements that are found across trials and see how well, how closely does the, does the movement on individual trials resemble this target movement, what we call template here? In most trials, the movement is very similar. So this is an ex parte animal that is able to reproduce this movement reliably. And there are small fraction of trials where the movement becomes dissimilar from the target movement. These are experts. They still make mistakes, right? The optogenetics stimulation induced movements or less consistent. Here. Grand mean shown in thick lines and individual animals are shown in thin lines. So the really good learned movements are induced in something a little more than half of the trials of the optogenetics stimulation. Any other trials, the movement is still induced, but not quite the clean target movement. So there's a variability even though we're doing the exact same stimulation, right? But the resulting movement is variable from trial to trial. So we wanted to understand this. Is that movement at the leper or this is the movement of delivery. Trial average. Yeah. We're doing some analysis with the video as well. We're recording video that we think the liver movement is captured as a large fraction of the barrio and sudden movements that things correlate with, yeah, hi, sorry, sorry. This might be embedded in one of the earlier questions. But when you activate those 20 neurons, are they all activated simultaneously or because they had some variability naturally? And I'm wondering if if you activate them in a way that was consistent with the original activation, would it be more behavior be more consistent? Yeah, I totally agree that that would be a prediction. We don't know. So right now we're activating all of them as simultaneously as possible. And it's just a short pulse of activation. We haven't played with the pattern, temporal pattern of that. Yeah. Good question. Okay. So something is variable from trial to trial, although we're doing delivering the same light stimulation. So we wanted to understand this variable d by looking at the population activity that is induced by this artificial stimulation. So we're activating neurons. At the same time, we're imaging all the new neighborhood neurons that were not directly activating. And many of them are driven by this artificial stimulation, indirectly, presumably through synaptic connections. And so here's a population activity pattern. This is actually from a voluntary trials, average lacrosse trials. And there are a lot of neurons high-dimensional. We reduce it using simple principle component analysis. And this, we can reduce this populace and activity to a trajectory in the high-dimensional space or reduced dimensional space. Starting from before the movement here in an open circle, and the movement starts right here. And then the population activity transitions along this trajectory. And what happens when we're activating neurons. So here's an average of the genetic stimulation trials. So before the stimulation, the population might be around here. And then we stimulate right here. And then the population activity follows a trajectory in this example, very similar to this unnatural trajectory bond three trajectory. And we find that this variability across animals of how well the learned movements is induced correlates very tightly with how well can we approximate the learned activity pattern. Here. The x-axis is how close these two trajectories are. The natural trajectory and the distance between the natural trajectory and genetically induced to trajectory that's in x-axis. So closer these population activity trajectories are the more similar than induced movements across trials. So it seems that this movement variability, variability of induced movements seems to be related to the variability of induced population activity. And this seems, seems so even on a single trial level. So here's a single trial example trajectory. And actually this is a trial average, but yeah, this is a natural trajectory. And now I'm going to show you three different trajectories of different groups of trials. Here is a group of trials where the Induced Movement is very similar to the template. And these are different groups of trials from the same animals where the induced movements are not very similar to the template learned movement. And you can see that when the resulting movement is very similar to the learner movement, the trajectory is also similar. But when the movements or dissimilarity to learn that movement, the trajectories are also these similar. And this is a very high correlation on the trial by trial basis. Very significant correlation. So it appears that even though we're applying the same stimulation to the same 20 neurons, the resulting population activity is variable from trial to trial and sexual deception. And that correlates with how closely we can reproduce the movement. But why is that? Why is the induced activity different even though we're applying the exact same stimulation. But this figure shows the clue to that, right? So the optogenetics stimulation is applied at these filled circles. All right? But even before this stimulation is applied, the populations at different places and in this application activity space. In this quote, unquote, good trials, where the movement can approximate to learn to move man, the pre stimulations population activity is very close to the voluntary trials. But when this pre, pre stimulation state is far from this voluntary state, the induced activity is also very different from the learned movement. And so we analyze this, the pre-activity, a pretty stimulation state. And we also see very tight relationship between how close the pre stimulation state is to the template pretty three-movement state. That has a strong predictive information about how well the stimulation can induce the learner movements. So we're not controlling this pretty stimulation state, different trials. The mouse is just in a different state. The M1 activities in different states. But when the stimulation is applied, when that happens to be when the mouse is, m1 is in the, say, pre movements, right? And then this 20 cell stimulation is sufficient to generate the naturalistic activity trajectory. And very nice, learn the movement. But when the exists, a pre-existing state is far from this good preparatory stage. Let's say there are 20. Self-stimulation is not sufficient to resemble the learned behavior. Yeah. So this is if you're computing your principal components from all 100 neurons in this case. So if you restrict it to the neurons that you're stimulating, did you get, Is it the same story? So is it the preparatory state of the entire network or is it just that the cells that you're stimulating quiet are primed and ready to go. Okay. Interesting. Okay. Yeah, good question. So I realize I didn't specify that for all of this analysis, we're not including the cells of our stimulating. And we have a separate set of analysis that shows that direct stimulation is very reliable. Across trials, we don't see a clear variability, at least that correlates with the variability of the induced movements. So every time we're activating the 20 cells very reliably. But the response of other cells to that stimuli is Sean is our variable. In the pre period is, I guess it's not necessarily super correlated between the AD neurons that he used for this and the 20 rounds but he stimulated. So there's a lot of the variants explained by the three stem movement and like EC2 here. Yeah. Well, I'm not sure I'm not sure I understand your question. There. Can you can say a little more. Yeah. So I guess what I'm wondering is what PC2, indeed, these neurons are activated during movement. So in this case, you're saying that variability in the 20 neurons are stimulated in the pre period does not explain much variability. But in these other neurons. So the principal components, I'm just curious. Difference between those two populations. Yeah, Sorry, that's not what I said. What I said is that the response to optogenetics stimulation over the target cells is very reliable. So there's not much variability. We haven't done the analysis to say which PCs corresponds to which kind of a variable t with sales, etc. Yeah. I was wondering if for your non outer genetic trials, the ones that are just the tone Induced Movement or the long 30 trial and motivate voluntary trials. Whether or not this kinda like pre tone activity then also correlates to the induced activity in the movement. Or if there's something about the tone that actually is setting this, this day. Yeah. Thank you for that. The last part of your question is really interesting, and I don't know if we, if I can address that, but that's exactly Actually the next animation. The first part of your question, at least. So, yeah, so the question here is, is this something really unique to observe and optogenetics stimulation? Just somehow the state responds differently than the population responds differently to the same optogenetics stimulation? Or are we actually uncovering something that we see in the natural behavior, right? So that drove us to do analogous analysis to see if the free movement activity state can predict the voluntary movement quality. And so that's the last data slide, data, data that I wanted to share with you today? And the short answer is it does. So we can look at where the population activity is during the, during the stage, right before the cue comes on. And we can do a fairly good job predicting how well the how good the trial is going to be. Right? So if the if the social there is a specific name, maybe I can go to the schematic. So this is what we think is happening here. So there is a preferred preparatory state of the population activity. And often in most trials can enter the state in experts. And at that point, the involuntary trials, it generates a reproducible trajectory that is reliably paired with the learner movement. If the network state happens to be in this state, activation of 20 M on layer 23 cells is sufficient to mimic the population activity. But this preparatory stage can be variable. And in some trials you're over here. And then whether you are stimulating them artificially or the mouse is induced move by sound cue. The resulting activity is not what the population is trying to achieve. And then induced movement is also very different. It's the same thing if you happen to be over here, you're in the wrong state. The stimulus on or Q will result in different activity patterns leading to different movements. So there's something that is really setting the preparatory state so that the population is really ready to drive this learned movement activity pattern. And that's when even in natural environment are in unnatural trials, the modesty is able to generate the right movement. And in those trials we can stimulate 20th cells and generate the same movement essentially. Yeah. Yeah. Yeah. Yeah. What do I mean by state? So it's just a location in the population activity space. This is all based on single-cell activity from on average about 400 neurons. And so basically it's a 400 dimensional space, right? Yeah, every cell has an activity level and it's up 100 dimensional space. We reduce it does to reduce noise, but that's not very important. So basically, so that's basically which neurons are active at the time. All right, so that's what we mean by state. So when the right neurons are slightly active, that's the right preparatory step. So it seems like the measurement of movement, its similarity to the template, like as you show, is like the distance to the template. So have you looked at other aspect of the kinematics like speed, acceleration or the extent of the movement like Really how much it overshoot or undershoot, and how, whether that would be related to what it turned off state and the optogenetics stimulation. Yeah, excellent question. We need to do a lot more and this is really new and we haven't done a whole lot. One thing I can tell you is that up to genetically induced movements tend to be smaller in amplitude compared to voluntary trials. They don't push as far. So, yeah, that is not something we're mimicking quite well and we don't know why. But there are other aspects that we should look at as well. I'm Chris rogers asks, is the pre movement activity state changing slowly, perhaps reflecting arousal or motivation? Or is it changing rapidly perhaps reflecting movement preparation. And could neuromodulators, modulators set the preparatory state? Yeah, yeah. Although these are all the questions that we think about, we don't have a really good answer to it. How quickly are they changing? We don't know. We don't know. We only did trial by trial analysis. Within trials, can they rapidly enter the right preparatory stage? We don't know. We don't know. I imagine install it, but we don't know. As a neuromodulator driven, very likely. Although it has, we think it has to be a high dimensional signal, or at least it's not a single dimensional signal because you must be able to prepare different movements, right? So it's a little hard to imagine that single neuromodulatory is sufficient for that. But yeah, we don't know. This is all speculation. Yeah, great question. We're getting that we're really interested in the function of noradrenalin in this aspect as well. Because there's a really nice paper from another group that showed that the noradrenalin can really flip the state between quiescence and movements in M1 and we're interested in looking at bat. So it seems like a component of learning might also be getting in the right preparatory state. So in the first part, you showed the learning is really about the learned movement. But I'm wondering if you've thought about wording of the preparatory stage. Absolutely. I agree. I agree with you. I think well, I think it's both, right. So even if we can somehow force the system to be in the right preparatory state in naive animal, I don't think stimulation can generate the whole activity. So this whole activity trajectory, I think, is also learned by synaptic plasticity. So I think both have to be learned, right? Yeah, I know less about how they're learning the preparatory stage. Even less about, I should say. We just have a clarification question. I might have just missed it from your last slide. But so in the naturalistic voluntary movement, like did you see the trend that the, the, the neural activity was like neural activity was set at the correct initial state most of the time or they're more likely to be. So, so this, this analysis. So basically where a lot of the trials lie in this regime where the preparatory state is very close to the target and then their movements are good. But in some trials there are far from that right target and then their movements are also bad. Alright, so just to summarize the second part of my talk, we think we showed that targeted stimulation of 20 movement related among neurons is sufficient to induce naturalistic movements. But that doesn't always happen very successfully. And variability of induced population activity reflects the variability of induced movements. And the pre stimulation population activity state now has a predictive information about what kind of hopeless and activity and movements are induced by optogenetics stimulation. Okay, So with that, I'd like to conclude by thanking all the lab members. The first part of the talk was a post-doc work of Nathan Heydrich, very talented individual who will be on the job market this year. So keep an eye on him. And the second part is run by onward, another postdoc, extremely one of the best experimentalist I've ever seen that makes very difficult experiments easy. So I can emphasize how hard it is that the experiment that we showed at the beginning. But yeah, I'm makes it look easy. Like to thank all the other members and thank you all for your attention. Thank you. Happy to keep answering questions, but I don't want to keep everyone from your next engagement either. Is there any question from the audience? One question on Zoom from Dieter Yeager, who's going to try to unmute yourself and ask a yes. Hello, I'm I audible? Yes, you are. It's kinda soft. Okay. So wonderful talk. Thank you very much. That was great. I have a question toward the first part. Since there's dopamine innervation of motor cortex, I was wondering whether this time plasticity was dependent on dopamine release. Yeah, good question. We actually have looked at that and I know there's a bunch of studies that have looked at more dopamine's role in regulating motor cortex plasticity. In our hands we see very little dopaminergic axons in the mouse M1. So we didn't follow up on that. Yeah, I don't know why. Actually, I haven't seen a lot of convincing Guyana anatomical evidence that midbrain dopaminergic neurons project to M1. So there could be other sources of dopamine. And so, so we're looking more at other neuromodulators. For example, we recently had a paper showing that the cholinergic innervation of M1 is important for regulating inhibitory neurons downregulate excitatory plasticity. So there's that kind of neuromodulatory control of learning induced plasticity in M1. Right? Thanks a lot. Yeah, I agree that the dopamine has been shown more on premotor cortex and ALM for instance. I guess that leads to the question how much the motor learning you think might also happen, say in N2 and higher-order motor cortices. Yeah, good question. I think it's distributed. One thing I can say is that when we block synaptic plasticity in M1, we can block learning. So it seems essential. But that doesn't mean that this is the only place where changes are taking place. I don't think it is. Thank you. Thank you. Thank you.