All right, good morning. It's such a pleasure to be here today. My name is K.M. Tai, and I am coming to you from the Salk Institute to tell you about the amygdala and circuits involved in valence processing. Some stimuli carry innate emotional valence. No training or learning is required to have emotional responses to them. Other stimuli, like the picture of these twin towers, do require associations to have an emotional response to them. For example, we would need to know about the events of 9-11 to be able to form the association between seeing this picture of the twin towers and the fear or anxiety or nostalgia that could be evoked from seeing them. In my field, it's been long debated about whether humans and animals can experience emotions in the same or different manner. Of course, it's impossible to determine whether the subjective experience of an animal's emotions are similar to my own, and I would argue it's also impossible to know whether my subjective experience's emotion is the same as another human's. So the big question is really, how do we assign motivational significance to sensory stimuli? So you take an individual, like Usain Bolt or a war veteran, present them with the same auditory stimulus, a gunshot, and in one case, you could form a positive association, thinking about all the gold medals, or in another case, you could form a negative association, thinking about all the death and trauma and anxiety and fear. In contrast, if you were presented with this auditory stimulus, a gunshot, many, many, many times with no outcome of any significance, then you would just be bored. You would habituate to this stimulus and it would not carry any motivational significance. So if we take this and think about the two-dimensional theory of emotion, we can plot on one axis intensity or arousal, or some might call it salience, and on the x-axis, we can plot valence or hedonic value, where on one side, there are very, very strongly negative emotional states like fear, anxiety, dread, sadness, or very positive ones, euphoria, excitement, joy. And think about this process. So, okay, let's think about the way that information needs to flow into the brain using the two-factor theory of emotion. Completely different concept. A stimulus is presented, and the first question that the brain needs to answer is, is this important? Do I need to attend to this stimulus? You might consider this to be the absolute value. What is really plotted on the y-axis over here? If it is not important and it's neutral, then you don't attend to it. If it is important, the next question is, is it good or bad? Is this something positive or negative? And if it's negative, we want to avoid it, and if it's positive, we want to approach it, generally speaking, you know, that's a simplistic view. So this is really the question that I'm interested in talking about right now. How do we tell if something is good or bad? Why is this important? Well, perturbations of motivational valence can be manifested in a number of different conditions. Many of them are mental health disorders. For example, an individual with too much motivation to avoid potential negative consequences in the environment and not enough motivation to seek potential rewards might be characterized with anxiety disorders. Conversely, if you had too much motivation to seek potential rewards and not enough motivation to avoid negative consequences, you might have a substance use disorder, for example. In another case, you might just have blunted motivation overall, perhaps a bias towards negative affect, and this might be manifested as depression. And this is very simplistic. This is just to show that the perturbations of motivational valence can be manifested in many different ways, and actually, any sort of abnormality or impairment in valence processing really is a hallmark feature of many different mental health disorders. Okay, so how does this work? How does the brain implement this? And so, actually, a lot of work, a lot of different research has investigated circuits that can underlie approach positive valence and avoidance negative valence signals. And so, there's the very simple, simple, almost a straw man model of labeled lines, which is very useful in sensory systems, but in terms of valence processing, when would you want this? So, maybe labeled lines would be useful in the context of valence for something like gustatory inputs, which are hardwired, something that's bitter. The first thing that happens, you just wanna plug in to a reaction or something that's sweet signals. There's rich caloric value here. And labeled lines just means that things that are relayed forward without much processing. So, the benefit of this is that it would be rapid and robust. The downside of a labeled lines motif is that it would not allow for flexibility, reversals, and learning other things. For example, how is it that we like bitter foods and drinks like coffee or alcohol? It's because we've learned to associate these flavors with the experience of taking these substances in. So, other motifs that are perhaps what I'm gonna focus on for today are, one, divergent paths, where different components that drive either approach or avoidance emerge from the same place where sensory input is coming in and could be fed down either pathway. This is work that I'll talk about today, led by Praneeth Amburi and Anna Baylor. Another motif, opposing components, can be seen when there are multiple cell types that have different functions, but yet have the same anatomical connection, meaning they originate from point A and send projections down to point B, but they have distinct functions. For example, in the lateral hypothalamus, my former graduate student Edward Nia found that there are both glutamatergic and GABAergic components of the projection from the lateral hypothalamus to the ventral tegmental area, and they have opposing functions and driving behavior. Fourth, really perhaps the most elusive, but potentially the most important motif is that of neuromodulatory gain. Different concentrations of neuromodulatory signal can have, in many cases, an inverted U-shaped-like curve or just a dose-dependent curve in terms of what cells they'll target and the effects that they'll have. And so one example of this can be seen from the work of Caitlin VanderWeel and Cody Siciliano in the prefrontal cortex. But today I'm gonna be focusing on the amygdala, and I will be focusing on examples of how divergent paths exist in the amygdala, and actually neuromodulatory gain can play a role in not only extending timescales to make it possible for learning to occur, but also to guide information down the correct path. Okay, so the outline for today is number one, where do circuits encoding positive and negative valence diverge? Two, what are the local interactions between these functionally distinct circuits? What happens in the amygdala with these neurons that project different downstream targets, and how do they interact, and how might those interactions change depending on internal state? And then finally, three, how does neuromodulation influence valence assignment? And then I'll close with an overview and outlook. So we have known for a very long time, nearly over 100 years if you consider it more broadly, that the temporal lobe, and then more specifically the amygdala as of 1956, that the amygdala is important for emotional processing of environmental stimuli. A monkey with amygdala lesions will lose his natural fear of snakes, for example. Human patients with bilateral amygdala damage will also lose the fear to snakes and spiders, have an impaired ability to recognize emotion in other people's faces, but in response to something like suffocation, will still display autonomic responses of a panic-like physiological response. And so this suggests that the emotional appraisal and the physiological or autonomic response are separable. In my laboratory, we work on mice, and if you take a coronal section of a mouse brain, you can find the amygdala here, which is an almond-shaped structure, and we can break that down. There are 13 sub-nuclei of the amygdala, but I'm only gonna focus on a couple of them today. The basolateral amygdala complex, which includes some sub-nuclei, as well as the central amygdala complex, which include the, I'll be talking about the central medial component of the central amygdala. So I might switch back and forth between those. So again, where do circuits encoding positive and negative balance diverge? Well, first, an introduction to the amygdala, which I doubt this audience needs, but I'll provide it in any case. The basolateral amygdala is thought to be cortical-like, because it's 90% glutamatergic pyramidal neurons, and it shows the associational processes that we would see in other cortical-like structures. The central amygdala, including its subdivisions, is largely striatal-like, 95% GABAergic medium spiny neurons, and so this is very much like the striatum, like a primitive striatum. There's a lot of support for the notion that the basolateral amygdala is a great site for divergence, to look for diverging circuits. Cellular resolution recordings within the basolateral amygdala, from a number of studies, have demonstrated that there are neurons that will have positive valence encoding properties or negative valence encoding properties, and in some case, both. Also, sensory information of all modalities basically converges into the basolateral amygdala, enabling this dual plasticity between a conditioned stimulus, for example, an auditory tone, as well as a gustatory unconditioned stimulus, or a punishing unconditioned stimulus, such as a foot shock. We also know that learning induces plasticity within the lateral amygdala, or the basolateral amygdala complex. And importantly, it occurs in this specific synapse. So I'll take a brief moment to make a didactic digression. Here, I just want to point out, this will be important for understanding a number of other data I'm about to show you, but basically, amphetamine ratio is a proxy for glutamatergic synaptic strength. So after long-term potentiation, or LTP, often you'll see an increase in AMPA receptor-mediated occurrence due to either an insertion of AMPA receptors or phosphorylation. In contrast, long-term depression reduces AMPA receptor-mediated occurrence, and then NMDA receptor-mediated occurrence are often thought to remain the same. And so after LTP, you will see an increase in the AMPA to NMDA ratio. Indeed, several groups have shown that fear conditioning can increase AMPA to NMDA ratio in putative thalamoamygdala synapses, right here. And similarly, as a graduate student in 2008, I also found that AMPA to NMDA ratio increases in putative thalamoamygdala synapses following reward learning. And so when I first found these data as a graduate student, I took it to the first meeting I ever went to, and a colleague sort of accosted me in the corner and just said, this makes no sense. How can the same mechanism underlie fear and reward conditioning? And this is an excellent question. So there are a couple of possibilities. One possibility is that, you know, it doesn't. Maybe the amygdala just encodes salience. And so what we're really looking at is individuals testing, you know, either fear conditioning or reward conditioning, but really anything salient learning would produce plasticity in the amygdala, that's possible. Another possibility is that maybe the amygdala is in fact the site of valence assignment, and this occurs via distinct downstream projections, which might be intermingled within the amygdala, so that's difficult to see without actually testing it. We know already that the central medial nucleus of the central amygdala is critical for the expression of fear. We've seen that optogenetically, you know, this is also true with chemical and electrical stimulation, but stimulating CEM neurons evokes freezing responses. Disconnecting the basolateral amygdala and the central medial nucleus abolishes fear expression. But of course, there are many diverse functions of the central medial nucleus, and it is not a homogenous structure with just a single function. It is a diverse structure with many different cell types that can produce a variety of different behavioral outputs as well. We also know, though, that the basolateral amygdala projects to the nucleus accumbens, and the nucleus accumbens is possibly the best, is certainly best known for reward-related processes, and we know that optogenetically stimulating basolateral amygdala terminals in the nucleus accumbens supports self-stimulation, intracranial self-stimulation, and real-time place preference. So what is the circuit mechanism for assigning positive or negative valence? So when I first opened my lab up at MIT in 2012, Praneeth Namvuri and Anna Baylor came hungry and ready to tackle this very big question. Anna Baylor is now a professor at Bordeaux, and, you know, she's fantastic, so you should check her out, and Praneeth is at Columbia. So the hypothesis is that basolateral amygdala neuron projection targets predicts learning-induced synaptic plasticity, specifically that if BLA neurons are projected to the central medial nucleus, if we pair a tone with a foot shock, that we will get LTP into these, you know, auditory inputs coming onto this central medial nucleus, where there's a convergence of US signals carrying information about the foot shock. In contrast, if we play a same tone and pair it instead with a reward such as sucrose, then we will see an increase of synaptic strength onto these auditory inputs after reward learning. So what we did, very simple experiment, was we just labeled these neurons. We injected retrogradely-traveling fluorescent beads into either the nucleus recumbens or the central medial nucleus. We then took separate groups of animals and either trained them to associate a tone with a shock, or they underwent reward conditioning, where we pair a tone with sucrose. And then we just recorded and stimulated putative thalamic inputs. So what we found was that synapses onto BLA-CEM neurons undergo long-term protection, a strengthening of synapses, as read out by amphetamine ratio, after fear conditioning relative to unpaired controls. Unpaired controls were also exposed to the chamber, exposed to the same number of shocks and tones, but they were explicitly unpaired in time. In contrast, after reward conditioning, we saw a reduction in amphetamine ratio indicative of LTG after reward learning in these synapses coming onto BLA-CEM neurons. Conversely, if we look at synapses onto BLA-NAC neurons, they undergo a reduction of amphetamine ratio. And after... Sorry, there's a little typo here. That's supposed to say fear conditioning. And after reward conditioning, you'll see a potentiation, strengthening of synapses here. And so one thing I want to point out, though, is that these naive groups are different. The naive group for fear conditioning and the naive group for the reward conditioning task are different only in one way, in that there's food restriction. So I won't talk about that too much this moment, but I just want to point that out, that the baselines are different because we food-restrict animals before reward conditioning to increase the motivation so that they're more likely to learn the task in a single day. So there are opposite changes in synaptic strength that occur after fear and reward conditioning. And there's a double dissociation. There's this opposite synaptic changes that occur after fear and reward conditioning in these functionally opposed projections. But is there a causal relationship? So we can photostimulate these neurons, and indeed, BLA-NAC projectors support positive reinforcement, intracranial self-stimulation, and BLA-CEM neurons support punishment. Animals will avoid the side of the chamber that is paired with photostimulation. So another didactic digression, just very quickly. If it's true that the learning that occurs here is occurring by your classic garden-variety NMDA-receptor-dependent LTP, what needs to happen is that glutamate is released from the presynaptic terminal. It binds to both AMPA receptors and NMDA receptors, but NMDA receptors is a magnesium blockade sitting in the pore. And what we need is depolarization of AMPA receptors to open, allowing this depolarization to push the magnesium blockade out of the NMDA receptor pore so that we can get calcium influx through here that is important for the cascade of events for LTP. So if you were to hyperpolarize these neurons, that should theoretically block learning. So what we did was we photo-inhibited either BLA to NAC neurons or BLA to CEM neurons bilaterally during either fear or reward condition. And what we found was that photo-inhibition of BLA to CEM significantly impairs fear learning and enhances reward learning. So that's kind of what I expected, but what I didn't expect, I sort of expected a different outcome here for the reward. I was expecting to see an impairment in the BLA to NAC productors, not necessarily an enhancement of CEM productors. And so this was very interesting. Inhibiting BLA to CEM predictors didn't really seem to change anything, whereas inhibiting BLA to CEM predictors either impaired fear learning or enhanced reward learning. So why are these, you know, threat or this negative valence processing circuit so dominant? So I'll come back to that. For the interim summary, I've just showed you that opposite synaptic changes map onto projection. After conditioning, we get synaptic strengthening onto CEM neurons. After reward learning, we get strengthening of inputs coming onto BLA to accumbens neurons. And activation of either projection either produces avoidance or approach. And then interestingly, there's this asymmetry here. The inhibition of CEM predictors impairs fear, but enhances reward learning. So this sounds great, but is it really that simple? And so to make a long story short, I'll just tell you that it's not that simple. So in this review article that I coauthored with my PhD advisor, Patricia Janik, we explored this possibility. And we're excited that, you know, we've got these new tools that help us to target specific cell types with specific projections. But really, even though we have this greater targeting specificity, it's still unlikely that all of the neurons that project from one region to another have identical functions and that any subset of cells that we target are homogenous. Even though we can stimulate a given circuit and get a behavioral output, it doesn't necessarily mean that's what all the neurons that we stimulated do. That would be like concluding that, you know, everybody in America agrees on who the president should be because we have one president. Clearly that's not the case. What we're looking at is we might still be observing a majority vote for a given behavioral readout when manipulating any circuit component. And there could be a lot of heterogeneity within that. More simply put, Eve Marder has said that optogenetic tools tell us what neurons can do, not what neurons do do. So we want to record from individual neurons and see what they actually do do. So the minimal criteria for saying that a neuron is valence encoding is, if you look at this, there's more granularity here, but I'm told that this is the maximum complexity that people can process in a talk. So on one axis, we have the response to the reward predictive condition stimulus. On the other axis, we have the response to the aversive condition stimulus, the shock predicting condition stimulus or whatever. What we're looking for are the different responses of the same single cell. So to be considered valence encoding, first, the cell needs to be task responsive. Second, it needs to have differential responding. If you get the same response to these two different cues, then that could just be a salience signal. And third, the responses need to be independent of stimulus features and be tracking the motivational significance of the stimulus. OK, so we basically wanted to explore this, explore the functional heterogeneity of the BLA, and we recorded in animals trained in a task where one cue predicted sucrose and another cue predicted quinine, a very bitter taste delivered into the mouth. And here I'm showing you the 1600 plus neurons that we recorded from. Anna Baylor here collected these data and what you can see, the Z-scored heat maps of each cell type where black is an inhibition and yellow is an excitation. And so what we found in a nutshell, if we look at all the data, is very consistent to what others have found. There are some neurons that selectively encode the rewarding cues, some neurons that selectively encode the aversive cue, some that encode both. And most, you know, about half of the neurons don't respond at all to either stimulus. But we didn't, you know, we didn't record from 1600 neurons just to replicate existing data, although it's always great to do so, and so we're glad that we did that. But what we did want to do is to be able to overlay structure and function. So using a well-established technique first developed by Susanna Lima and Tony Zader's lab. Called photosimulation-assisted identification of neuronal populations, which I will call phototagging, what we do is implant an electrode attached to an optical fiber and then express channelrhodopsin in cells of interest. In this case, projection to find basolateral amygdala neurons and then record from all the amygdala neurons near the probe. So if neuron A spikes, we record that spike. If neuron B spikes, we record that spike. And then only after the behavior, after the animals completed the task, at the very end of the session, we can deliver light pulses and identify the neurons that are photoresponsive versus the ones that are not. So importantly, a very, very, very critical caveat for me to point out is that of recurrent excitation. And this is important in any brain region where you're trying to do phototagging, where there could be recurrent excitation. Hippocampus, cortex, amygdala, these are all regions known to have recurrent excitation. And what I mean by that is. You know, when we're doing this experiment, ShineLight, it's easy to tell the difference between CHR2-expressing neurons and their non-expressing neighbors that don't fire at all. It is much harder to tell the difference between CHR2-expressing neurons and non-expressing neighbors that do receive recurrent excitation from these CHR2-expressing neurons. So we don't want to mix these gray neurons that might still spike in response to light from the ones that are actually the CHR2-expressing ones, since those are the ones we're trying to identify. So. One way around this is that we can record from individual neurons ex vivo where we can unequivocally identify whether they're expressing CHR2, which is typically fused, in this case, it's fused to a yellow fluorescent protein, as well as patch onto neurons and get a very classical, you know, peak and steady state response to a steady light pulse. And so what we were looking for on the on the X axis, we just have the number of units and on the Y axis, we have the photoresponse threshold, the latency of that it takes from when that light comes on to when they spike. Cells, you know, for the non-expressing neighbors, there is a synapse between. And so sometimes, oftentimes, in fact, you can create separate distributions that don't overlap, where CHR2-expressing neurons have fast responses to light, short latency thresholds, and then short latency responses, and non-expressing neighbors have slower ones. And so when it's separable, that's great. It's easy. Sometimes, however, they're partially overlapping. And then there's there's the challenge of determining which is, you know, false negatives and false positives when when there is this overlap. And sometimes the overlap is so egregious that just phototagging is not a suitable approach for the particular preparation. OK, so I won't go through all the details of the data that we got, but I'll just summarize to say that basolateral amygdala to NAC predictors predominantly encode positive valence. Here, I'm just showing you the Z-score of the response to the Sucrose predictive Q and the difference relative difference score for the Z-score of the condition stimulus predicting quinine. So BLA to NAC predictors predominantly encode positive valence, BLA to central amygdala predictors predominantly encode negative valence. And then the eventual hippocampal predictors did not have a significant bias for either CS using these multiple different approaches. And so we analyze this in a number of different ways. And if you're interested, you can see all these approaches in our paper. OK, so that should bring us to the second part of this talk. So I've already shown you now that there are these distinct circuits that encode positive and negative valence. This can be seen in the synaptic changes that occur in the causal manipulations we perform on these different circuits and from the recordings that we do at a resolution when we photo identify neurons that go to a given projection. But what are these local interactions that occur between these functionally distinct circuits? These neurons are all salt and pepper kind of mixed in with each other. So what gives? And why would these neurons be intermingled? Is there a purpose for this? Why would you want a salt and pepper? Why would you want that there to be integration of these functionally distinct neurons? Perhaps this would facilitate circuit motifs like mutual inhibition. You know, if you're running away from a predator, you can't think about food. And if you're trying to mate, you need to focus on that, whatever. So you don't want to be trapped between two conflicting. Actions. It's also possible, and I only when we wrote this review, I only put this in because it's technically another possibility where there could be a unidirectional inhibition that that produces an asymmetry. Or there could be a neutral or excitatory relationship. So I apologize for changing the color scheme. The neurons typically in this talk shown red are now shown blue. And this is just a movie created by Anna Baylor and Craig Wiles with the help of Kwan-Hoon Chung at MIT. And you can see that BLA to NAC neurons and BLA to central medial neurons are, while there are gradients and hotspots, they are largely intermingled. We've done this in a more quantitative fashion in our 2018 cell report paper, another paper led by Anna Baylor, looking at the anatomical distribution of projection-defined basolateral amygdala neurons. Importantly, coming back to this asymmetry concept, and remember the data that I showed you, that BLA to CM, you know, this is CM, CEA, CM is part of the CEA. These neurons have more ability to control behavior. To control behavior than do BLA NAC predictors. And we see this borne out again when we record locally in vivo. Of course, there are the phototype neurons. Those are the ones we're looking for. They're also the non-expressing neighbors, the photo excited neurons with a longer photo response latency. And then there's this huge group of neurons that are photo-inhibited, that are silenced by the activation of the CTR2-expressive neuron. And look at these numbers. You know, the number of neurons that are silenced when these central amygdala projecting BLA neurons are activated is massive. If we quantify this and compare to other projectors normalizing for CTR2 expression, you can see that indeed, here I'm showing you, you know, CTR2 expressors, neurons that are excited by those projectors, and neurons that are inhibited by those projectors. And you can see that the BLA to central amygdala neurons have greater influence over their neighbors. They have the greater ability to suppress the activity of other basolateral amygdala neurons, perhaps explaining why BLA to central amygdala neurons, when inhibited, can both impair fear conditioning and enhance reward conditioning, whereas silencing BLA to NAC projectors doesn't do much of anything. It's because silencing these central amygdala projectors relieves this huge inhibitory suppression to neighboring neurons. So, why might the brain work this way? Why would there be an asymmetric or unidirectional relationship? And I'll just speculate and say, well, you know, animals, when they engage in reward seeking, be it socializing and mating, eating or drinking water, especially if you think about mice in a burrow, because that's what we're studying. I study mice. I don't, I don't study zebras as misleading as this slide may be. Mice, to seek rewards, have to venture out of the safety of their burrow. So, perhaps reward seeking is inherently risky and priming escape is a good insurance policy. You know, it's much more important for survival that you live to see another day. You can eat or drink or mate a little later. Right now, you need to escape this predator or else you won't get the opportunity to do any of those other things. So, perhaps it makes sense that there would be an asymmetric or unidirectional relationship. So, perhaps it makes sense that there would be an asymmetric relationship because survival and escaping a predator is paramount to any other needs. OK, so the new data that I really want to talk about and the most of what I'm about to show you is unpublished, is how does neuromodulation influence valence assignment? How does the brain solve the quote unquote valence assignment problem? And what I mean by that is how do the synapses know which postsynaptic neurons are which? And this is also a question of timescales as well. So, if we think about, you know, Hebbian spike timing dependent plasticity, you know, you get you have a window for LTD, you have a window for LTP. And really, this window is pretty short. You know, we're talking about tens of milliseconds that would shift the difference. The timing is very precise. If we're even out, you know, 100 milliseconds, it doesn't really work anymore. Similarly, if you look at this is the basic activity of a basolateral amygdala neuron recorded in vivo on the first day of learning or the or the third day of learning. And you can see that this this response sharpens with learning. However, we are back down to baseline levels of firing at about 100 milliseconds, certainly by, you know, a couple hundred milliseconds. But when we think about most tasks, both fear conditioning and reward conditioning, the condition stimulus is quite long. The onset of the condition stimulus. See here, we're playing a longer tone, but you really only see an onset response spikes that are occurring with the onset. And then, you know, they go back to basal levels for a long time. So 10 to 20 seconds later, the unconditioned stimulus is presented. And this is common. We are able to learn the associations between temporal events that occur on this on this timescale easily. However, it's not clear how we do this based on Hebbian Hebbian plasticity alone. This timescale doesn't, you know, it doesn't actually make sense. You know, neurons aren't super active at the time that the unconditioned stimulus is presented again, except for in response to the shock. You know, the neurons that are being driven by this, this condition stimulus might not be active anymore. So how does this work? How does this all add up? The timescales that we're talking about are orders of magnitude off of what those that should matter for Hebbian plasticity. So what I want to say here is that on some level, we think about neuromodulatory systems, because that is that is a potential strategy that the brain could apply to extend the timescales to to form associative memories. And then again, there's this question of how do the synapses know which postsynaptic neurons or which? How do they know the downstream projection target or or other features of these of these neurons? And so one thing we were thinking about would be maybe there's something on the surface of the cell of, let's say, for example, BLA to NAC predictors versus BLA to CEM predictors that tells the presynaptic neurons or terminals their identity in some way, something like a name tag, perhaps a surface receptor. So we performed transcriptomic profiling in collaboration with Jesse Gray and Suzoku Urozu, and we found that there are a number of differentially expressed candidate genes in these projection defined populations of BLA neurons. And one of them is the neurotensin one receptor, which we can validate in terms of relative expression is indeed different between NAC predictors and CEM predictors. So why did we single out the NTSR1 receptor? You know, there's a lot of different candidate genes that we came upon in this transcriptomic profiling, but there's a few reasons. NTSR1 controls expression of a GPCR and GPCRs in general. So this could be many different types of receptors, but GPCRs have the capacity to extend timescales. The function of NTSR1 receptor has been linked to LDP and the BLA, as well as contextual fear conditioning. And then for me, most significantly, Kimberly Kempadu showed that neurotensin has been shown to have differential effects when applied at various concentrations on the BTA dopamine neuron. So this is a different brain region, but the key is that neurotensin can either facilitate glutamatergic transmission at some concentrations or suppress glutamatergic transmission. And so this at different doses. So if it's possible that the same, that different doses of neurotensin can have opposite effects, either facilitating glutamatergic transmission or suppressing glutamatergic transmission, then maybe the same concentration of neurotensin could be a signal to different populations of neurons with different expression levels of the receptor to guide information down one path or another. We explored this. We did an experiment where we put NTSR1 antagonist non-specifically. Mind you, this is everything in the basal lateral amygdala. This is just a pharmacological manipulation, and then train animals on either reward or fear conditioning. What we found is that for reward conditioning, animals will enter the port more and learn the task more quickly if their NTSR1 antagonist is on board, we're blocking signaling relative to vehicle. But in contrast, we see a slight non-significant trend towards an impairment or maybe no response with fear conditioning. But this is a valence-specific effect. Further, if we look at the glutamatergic facilitation or suppression, do we see dose-dependent changes that are different for these different populations of neurons since BLA-CM predictors have five times the expression level of NTSR1 receptors as BLA-NAC predictors do, so it makes sense that there would be some dose dependency and that it would be different between these populations. We do indeed see that. What we would consider a relatively physiological concentration of neurotensin, you can see that BLA-NAC predictors at 10 nanomolar of neurotensin show a facilitation of glutamatergic transmission blocked by the antagonist. Whereas in contrast, BLA-CM predictors show a suppression of glutamatergic transmission also blocked by the antagonist. This suggests that neurotensin and the basolateral amygdala is important. But where does neurotensin come from? Neurotensin is released by many different cells throughout the brain. But for the basolateral amygdala, what is the source of neurotensin? We identified a few different sources, the medial geniculate nucleus, the ventral pacampus, and the paraventricular thalamus. Although I don't have time to show you all the data, we had the most compelling preliminary data with the paraventricular thalamus and have focused on that. I also don't have time to show you these data, but when we performed ex vivo recordings and looked for glutamatergic co-release, we found that neurotensin-nergic neurons, at least for each of these three inputs, all of them showed co-release of glutamate. That's something important for thinking about our next experiments. That is a caveat to consider in this particular experiment. But first, we wanted to see, okay, what is photo-stimulating these neurotensin-nergic neurons in the paraventricular thalamus projecting to the basolateral amygdala? How does that change behavior? What we found was that this gain-of-function manipulation on the PBT to BLA pathway facilitated reward learning and impaired fear learning. Importantly, again, this is just one prediction. There's lots of different sources of neurotensin. But here we are also co-releasing glutamate. This was something to consider. Here, are we really looking at the effect of neurotensin per se, or might we be also looking at the effect of PBT to BLA glutamate? The following experiments were performed by Hao Li, a phenomenal postdoc in the lab, who's done experiments we were attempting for many years and one technical challenge after another, and Hao managed to overcome them all. He did my dream experiment, which would be to, number 1, compare groups where neurotensin is on board, versus control groups, versus experimental groups where we crisper out the neurotensin gene and prevent the production of neurotensin selectively only in this pathway. This is projection-specific, crisper-mediated knockdown of neurotensin, then recorded in both of these populations of neurons in control and crisper animals, and then also photo-tagged and photo-identified these BLA to NAC or BLA to central amygdala projections. Okay. First, just showing you here, we were able to use crisper to inactivate the neurotensin gene. As you can see here, the neurotensin mRNA is reduced, whereas Vglut2 is still there. In contrast, for the control group, neurotensin is still there, as is Vglut2. What we found here is that crisper inactivation of the neurotensin gene in the PVT to BLA impairs reward learning and promotes fear learning. Again, the opposite effect of what we found specifically in this projection. This is distinct from nonspecific pharmacological experiments. Then what happens? The answer is complicated. Recording from many hundreds of neurons, what we can do is plot each individual neuron as a trial-averaged heat map, looking at the basal response, and the phase of changes occurring when we present a condition stimulus that predicts reward, a condition stimulus that predicts a neutral cue that doesn't have an outcome, or a condition stimulus that predicts foot shock. What we found is that we can functionally cluster using functional hierarchical or glomerative clustering to functionally cluster these different populations of neurons in the BLA, and look at how for each cue, how the crisper and control groups compare. If you look broadly, what you might be able to pull away at a quick glance is that in almost every case, what the effect of crisper on the neural coding or the neural dynamics is to reduce them. The amplitude of all of these valence-specific responses are muted or blunted when we've used crisper to knock down the NT gene. If we look at BLA to NAC projectors, in our control group, we replicate this effect, that BLA to NAC projectors in general preferentially encode positive valence, and this effect is attenuated when we crisper out the neurotransient gene in the PVT to BLA pathway. In contrast, the BLA to central amygdala projection, again, we also reproduce our previous results, that these BLA to CAA neurons predominantly encode negative valence. Again, when we crisper out the NT gene, these neurons lose that negative valence coding. That's just shown here. You can see the distribution of the proportion of these different functional clusters in each of these projection-defined populations. Then if we look at the ensemble activity in the form of neural trajectories, basically thinking about the ensemble dynamics as very high-dimensional space. This is just an example to explain what I'm doing here, in dimensionality reduction, plotting of neural trajectories. If we want to envision what the entire ensemble is doing, say there's 685 neurons in our ensemble, we could theoretically plot the ensemble dynamics in 685 dimensional space. Obviously, that's very difficult to display, so we can use principal components to reduce the dimensionality and then plot instead the axes that account for the most covariance in the dataset. If you're not familiar with this, don't worry about it too much. Just think about it as an abstraction of what the whole ensemble is doing. For the control animals, we're starting here and we continue on this path, and then the cue comes on. In the case of shock cues, there's a trajectory that goes this direction. For neutral cues, there's a trajectory that goes in this direction, and for sucrose cues, it's going in a different direction. You can see how the neural dynamics are different depending on which cue is presented. Here are the trajectories for the CRISPR group. If I blow this up, since it's so small that you can't really see them, and what does it mean that they're so small? When the trajectory lengths are reduced, this just means that there are reduced ensemble dynamics. A longer trajectory would indicate faster changes or greater changes in the neural dynamics, so greater neural dynamics, and shorter trajectories indicate fewer neural dynamics. You can see a similar pattern is occurring here. You do see this separation of shock cues, neutral cues, and sucrose cues, a very similar yet impoverished structure of what we see in the control group still exists here, but notice the dramatic shortening of the neural trajectories. And then finally, if we were to try to decode what the behavior is, first, if you train with control data, data collected from control animals, and then test on distinct data set but also data from control animals, we get very, very, very, very high decoder performance. The amygdala is very, very good at determining, you know, whether it was a shock cue, a neutral cue, or a reward cue. However, if we look at the CRISPR group, and you can see that it still works. It's significant above chance if we train and test within the same group, but it's significantly worse. The accuracy is less, and that makes sense, right? You know, if you have these big, clear neural trajectories that representations are dynamic and distinct, it's gonna be easier to decode than if they're not. And some people might be wondering, why is the chance not 50%? And it's just because the number of trials that we present, we bias it towards more rewards just so the animals don't give up and quit on us. And so that's a task feature. Okay, now, if we train with control data and test on CRISPR animals, or train with CRISPR animals, data from CRISPR animals and test on control animals, we get basically chance performance. You don't do well at all if you're, you know, testing across groups. And this just suggests that the coding rules have changed from control animals to CRISPR animals. So the NT gene and neurotensin release from the PBG-BLA pathway, a very specific source of neurotensin, reduces the decoding accuracy of behavior from the basolateral amygdala. Okay, so finally, Anish Bal in my lab has applied computer vision tools for analyzing behavioral motifs with a little bit more granularity, not just freezing and port entry, which is, you know, what we normally use, but using deep lab cut to look at the XY coordinates of the head, nose, and tail, and just get a little bit more details here. So after some feature extraction, dimensionality reduction, and then watershed segmentation, Anish can break up reward trials and shock trials into smaller subtypes of behavior. So for reward trials, there are seven subclasses. For shock trials, there are five. Within each of them, you can see the density, but also what is this? So here, this just shows actually what the watershed segmentation has done with low motivation state, which we just say is it takes them a long time to get to the port versus the high motivation state. They're right on top of it. They are in the port as soon as the cue is playing. And you can see that there are some cases where the animals are getting there slower, and some cases where they're getting there faster, and similarly, or for the shock trials, animals can either freeze or dart. Typically, animals, freezing is considered to be a passive avoidance state versus darting is an active avoidance state. And so you can see, again, that in terms of the responses, there are different subclasses of freezing, and there are also different subclasses of darting. So this is important because if we're changing the behavior, we could also be changing the neural activity. So we wanna match up similar behaviors, and since the distribution could be shifted, we wanna control for that. So we can do that here and look at control versus CRISPR animals and what changed. So first, the neural response, the reward cue, has changed relative to controls. There's a lot of very dynamic changes in firing, and then smaller changes in firing with the CRISPR group. Similar to the shock group, we have overall lower neural responses to the shock CS. Interestingly, though, for the shock CS, the CRISPR group has more neural encoding for certain behavioral motifs that the control group had less changes in firing for. But again, to take away that the control groups have more dynamic encoding, and then when we CRISPR out the neurotensin gene in the PVT to BLA pathway, we reduce the amplitudes. And this is more prominent in the BLA to CA population. Okay, so just to summarize and close, what I've told you so far is that the basolateral amygdala is a divergent site for circuits encoding positive and negative valence, that these neurons do functionally locally interact both with excitation and inhibition, and that there's an asymmetry here within the basolateral amygdala, and that finally, neuromodulatory influence can have a profound impact on valence assignment within the basolateral amygdala, both at the learning level and at the expression level, as well as the neural ensemble dynamics representational level. Okay, so to close with my overview and outlook. So really, I think this is exciting because we can identify differentially expressed genes in what we've already determined to be functionally distinct circuits. And this approach of looking for differentially expressed genes in functionally distinct circuits can be leveraged for selective control of neuronal populations, circuit-based drug discovery. Potentially, we could find small molecules or other drug targets that would allow for more efficacious therapeutic effects and fewer side effects. Because we're being more specific, the hope is that there will be fewer side effects, and we can selectively screen for only the therapeutic effect. And so just to close this loop with the bigger conceptual framework, this neuromodulatory game motif is also represented within the basal lateral amygdala. And big picture, how do we get from where we are right now with current treatments that are ineffective and have a lot of side effects, to a future where there are effective, lasting treatments without side effects that work for everyone, and every individual has a mental health disorder treatment that works for them? Well, to get from where we are now to where we want to be, there are a lot of steps. First, we need to think about how we got to where we are now. Well, we have ineffective treatments that are not effective for everyone or have side effects because we've discovered them through nonspecific trial and error. This has occurred because maybe there's poor disease classification and poor understanding of comorbidity, and we're using large buckets of very heterogeneous symptoms. For example, in depression, you sleep too much or too little, eat too much or too little. These are very different behavioral phenotypes and they're all lumped together into one bucket. Perhaps there are many different underlying pathologies or etiologies that explain these behavioral symptoms, even if they're similar, and certainly if they're different. So we just generally have a poor understanding of how brain gives rise to behavior. That is the big problem. There's no getting around this. There's no shortcut, you know? So what we first need to do to build a better future is to identify specific neural targets, figure out what common pathways can explain comorbidly expressed mental health disorders, and develop neural circuit-based therapeutic development. Use what we know about neural circuits to discover new treatment strategies. Okay, and so today, I've talked to you about the amygdala. The amygdala is a remarkably compact, multifunctional region that has been conserved across evolution and has scarcely changed from 70 million years ago in species that are reptilian species all the way to the human brain. This is not true for many other brain structures like the cortex, but the amygdala is well-conserved and this has the dual benefit. From a basic science perspective, understanding something, these simple circuits can help us understand more complex ones. And from a therapeutic perspective, if we use mouse models to study a circuit that's well-conserved, there's a greater likelihood that there will be translational relevance for humans. So with that, I'll close and thank you for your attention. I thank everyone in my lab who's done this work. I tried to acknowledge the specific individuals along the way, our collaborators, Feng Zhang, Kerry Ressler, Ilhafeet, and those who have provided us with reagents, our funding sources, and you for your attention. Thank you.

amygdala

valence processing

emotional responses

two-dimensional theory of emotion

neuronal populations

conditioning

fear learning

neuromodulation

mental health disorders