Probing memory circuits in the primate brain: from single neurons to neural networks
Date Posted:
March 25, 2019
Date Recorded:
March 22, 2019
Speaker(s):
Julio Martinez-Trujillo
All Captioned Videos Brains, Minds and Machines Seminar Series
Description:
Julio Martinez-Trujillo, Robarts Researc & Brain and Mind Institute, University of Western Ontario
Abstract: The brain’s memory systems are like time machines for thought: they transport sensory experiences from the past to the present, to guide our current decisions and actions. Memories have been classified into long-term, stored for time intervals of days, months, or years, and short-term, stored for shorter intervals of seconds or minutes. There is a consensus that these two types of memories involve different brain systems and have different underlying mechanisms. In this talk I will present data from different experiments in non-human primates examining brain circuits and mechanisms of both short-term memory and long-term memory.
Biography: Julio Martinez-Trujillo is Professor in the Department of Physiology and Pharmacology and Scientist at the Robarts Research Institute. He holds an Academic Chair in Autism. Prior to joining Western University in 2014, he was Associate Professor in the Department of Physiology and Canada Research Chair in Neuroscience at McGill University.
DIEGO: Good afternoon, everybody. Welcome to another seminar hosted by the Center for Brains, Minds, and Machines. Today the center is happy to host Dr. Julio Martinez-Trujillo coming from Canada. So he's enjoying the weather, even though it looks really bad in here.
A little bit about him-- Julio did, first, his medical degree in Cuba, finished in 1991. And after residencies, ended up as an assistant professor in the Javeriana University in Bogota, Colombia. One interesting thing is, as I tell this, try to plot in your minds what the temperature of these places is and see the trend.
So going from Cuba down to-- up in the Andes Mountains. There, as assistant professor he got some experience in neurology, looking at some patients with certain deficits that involved attention. That's where things kind of went in that direction.
He realized that perhaps being a medical doctor had its limitations in the contributions you can do, and that, perhaps, as a researcher there are other things that he could contribute. And so he went on. He decided that he actually wanted to advance the field, rather than just directly treat patients.
And then he chose to then go to start his first master's and doctorate in the University of Tubingen under the supervision of Steffen Troie and Peter Tier, which he completed with a magna cum laude in 2000. Now, being in Germany, a little colder, he decided that's not cold enough for him. So he decided to head, now, to Canada.
And ended up doing his post-doctoral fellowship at York University with Dr. [INAUDIBLE]. And there he did, also, very exciting research on eye movements and the relationship to what he was doing in attention. Just to not be quite satisfied with the temperature, then he decided that he wanted to move from Toronto to Montreal.
Ended up in one of the most-- one of the coldest cosmopolitan cities in the world. And I think there he thought that was enough. At McGill, he became associate professor in the department of physiology. And he started his own lab there, which I had a great pleasure to be part of for several years.
I would almost say too many, but it was so productive, I can't. So it's, you know, Julio's close to my heart. I can say that he is beyond a researcher, a neurophysiologist. He takes care of his people beyond that he knows how to bring some of his personality into his people, to tell him how-- tell them about how to do research.
He cares about those things. He doesn't just want to get things published. He wants to discuss. He wants to get-- he wants passion in it. And that's something that gets you excited about doing research. And that's invaluable.
Then, as I said, he had enough cold, so he ended up heading, after several years, in 2014, to University of Western Ontario, or Western University now, where he is now full professor at the departments of physiology, pharmacology, and psychology.
And now moved his lab to Western and is running it also at the Roberts Research Institute and the Brain and Mind Institute. And there, his lab continues to thrive. And he has been able to introduce a whole lot of really interesting topics.
So now, the initial studies on attention that he did with Steffen Troie are now a very small portion of all the really exciting things he does. So he went on to studying memory, short-term and long-term memory circuits and learning, and eye movements, and autism. And I don't have enough breath to talk about them.
And I think I should let Julio tell you everything about it. So today's talk is titled probing memory circuits in the primate brain from single neurons to neural networks. So without further ado, please help me welcome Julio Martinez-Trujillo.
[APPLAUSE]
DR. JULIO MARTINEZ-TRUJILLO: So I really thank Diego for this generous introduction. Maybe I don't deserve it that much. But I want to say that I'm very happy to be here. I think that this is a great place to do science. And definitely thank you very much for the invitation. And all what I have seen today is just amazing.
I also want to apologize for a small misspelling in pharmacology. It's not pharmacolary, it's pharmacology. Oh, for the lighting?
DIEGO: Yes, better lighting.
DR. JULIO MARTINEZ-TRUJILLO: That sounds good. So to illustrate a little bit about the research we're talking about today, this is a small-- oh, yeah, that's that. This is a video game that some of my students in the lab program-- for fun. It's called Game of Runes.
So what you have is a subject that basically got into these-- supposed to be some kind of magic stone. So you pay close attention to that. And now there are two runes, that they're going to appear there. You're going to remember those. Then, you have to go through this corridor.
I mean, the graphics needs to be upgraded because that was on RealEngine 3. But now, I think that it should do better than that. Now, you choose those runes on this big cave of the droids. And then you get into this tomb. And if the runes are correct, you keep going. If the runes are incorrect, they dumped you into some kind of dungeon. I don't want to go into that.
So, basically, in this task, you need to pay attention to the runes. You need to keep in short-term memory some representation of these runes. And actually, you keep-- if you keep doing the task again and again, you're going to probably learn the runic alphabet, which is somewhere in northern Germany and Holland today.
I'm going to be talking about two aspects of these cognitive functions-- short-term memory and long-term memory offline storage. So I always say that a challenge to neuroscientists is to find out how are neurons interconnected in the brain, within and between brain regions, and how do they interact to produce these functions?
I always try to-- I'm going to show you these dots. And I'm going to start naming some of these dots, and you're going to tell me what that is-- probably, some of you will. Diego, don't say it. Anyone know what that is? So that's a picture of the Montreal subway.
So and I feel that pretty much where we're going, it's like if we go 3,000 years ahead, and we go to the city of Montreal, and we have to start digging and discovered how these people transported-- go from place A to point B.
What we're doing in the brain is trying to pump the brain and trying to do fMRI [INAUDIBLE] from the nodes of this network, and to do anatomical studies trying to find the connectivity between the nodes, and then trying to find out how many passengers are transported from one train station to another-- so at every given moment.
So knowing the structural connectivity, you don't know the functional connectivity or the function still. But the worst thing is that now, these nodes they don't look all alike. So in all the brain areas, they don't have the same anatomical structure.
So, in the talk, I'm going to talk about calling up short-term memory by single neurons and neural ensembles, mainly in the prefrontal cortex. Also, I might touch some other areas of the brain. The coding of long-term memory by single neurons and neural ensembles in the hippocampus.
So the macaque brain-- information goes into the LGN. And from there, it goes to V1, and going to two main pathways. And information also flows back into other-- into area. So there's many connections-- feedback connections and feed forward connection.
A definition of visual working memories-- the maintenance and manipulation of visual information relevant to behavior for short time intervals, generally seconds. I'm going to stick to this definition of working memory, mainly to the maintenance aspects of working memory, more than manipulation, which is what most electrophysiologists do-- working in non-human primates.
So several decades ago, Funahashi and Goldman-Rakic, actually-- they described those cells in the lateral prefrontal cortex of monkeys. That if you train a monkey to remember one of these different spatial locations, you see that during the period when you turn the targets off and the monkey's remembering the location of the target, you see this increasing spike in activities.
That's what you call a persistent activity or sustained activity. And this is the most accepted neural correlate of working memory, so far. So, however, people thought that this persistent activity was restricted to association cortices. But new findings in the fMRI literature found that people could decode, actually, from visual areas, the contents of working memory.
So that poses the question-- is, really, working memory something that is neural correlates found only in regional areas? Or this is something that is found, also, in several other brain areas, like, for example, sensory cortices? So we set in the quest to answer this question.
And that was Diego's project. And we talk about how we could map different brain areas with the same task and actually look at the neural correlates in working memory. So what we decided to do was to put different electrodes in different brain areas.
One was in MT, the other in MST, and the other in the lateral prefrontal cortex of macaque monkeys performing a working memory task. And to examine whether working memory actually exist-- the correlates of working memory exists in these three different areas. So, by the way, these areas actually project to each other forward and feedback.
So here is the results in one slide. I mean, I can't summarize all the work. But I have to move on because I want to show some of the data. So in area MT, this is actually the sensory response to this sample, where what the animal has to remember that was in motion direction.
And what you see here, in the delay period, would be the response during the working memory period. So it's to see if these cells actually responds during this delay period. And gradually, to these four different stimuli, where motion directions-- we say that these cells encode working memory-- or encode working memory for motion direction.
MT didn't show the pattern. But MST and the lateral prefrontal cortex-- they both show persistent activity that represented the contents of working memory. So, from this work, we conclude, actually, that working memory is presented in association cortices MST and area 8A 46 in the lateral prefrontal cortex.
I'm going to use a 8A 46 and lateral prefrontal cortex interchangeably here, in this talk. But it's not presented-- it's not in sensory cortices. So what we concluded was a transitioning cortical architecture in this different brain from sensory areas to MST that allow, somehow, a working memory representation to arise.
Now, the next question that we set to answer is-- what are the feature of cortical microcircuits that allow persistent activity to emerge? And we came with two very simple hypotheses. The first hypothesis was that persistent activity arises from intrinsic properties of neurons that emerge in association areas.
For example, the ion channels composition. If you hit a neuron with the stimulus, and you stop the stimulus, the neuron is going to keep firing beyond that. It's equivalent to a longtime constant. So it's a very simple model.
And the second model was a persistent activity arises from network properties, such as recurrent dynamics that happens between neurons in these different-- within a microcircuit. This is not that crazy. For a system neurophysiology that is used to see the brain as a network it looks a little bit extreme.
But if you look at-- of course, the third hypothesis wanted to-- if you look at several papers, actually, they have shown that persistent activity could exist in slices. Actually, here, in this paper, what they did in [INAUDIBLE] slices, they applied carbocoal. We're doing patch clamp studies.
So they patched the cells and applied carbocoal. They hit the cell with a stimulus. And the cell show persistent activity. In this paper that is even more interesting, what they did was they just-- recording from macaque lateral prefrontal-- infero temporal neurons-- they inject a stimulus into their neurons.
And they see persistent activity in the proportion of neurons, even when you isolate tight synaptic transmission in those slices. So that was very interesting for me. And also, they found that happens in human and monkey neocortex, but it doesn't happen in the rat neocortex, at least not if you don't put carbocoal.
So the first experiments that I'm going to show you-- we decided to go and do patch clamp in slices of lateral prefrontal cortex in the monkey. So we put a team together. And, actually, we developed a procedure in which we just take slices of prefrontal cortex, and we send it to the patch clampers. The patch clampers inject currents using the Allen Institute protocol.
And then we observe if, after injecting currents, we see persistent activity in those cells in the lateral prefrontal cortex. Here it is, just a summary of the procedure. This is an interoperatory image, where you have the arcuate sulcus and the principal sulcus. Here, you have the pieces of the brain. We have to put that very quickly into four degree saline and transport it.
I think that someone is doing that here with human tissue. You know, probably, what we are going through. And here, you have some pictures of the patch clamp. And here, basically, the experiment. You block synaptic transmission. You inject a current. And then you see if there is persistent activity in neurons after you interrupt the simulation.
The results were a little bit surprising. This is an example neuron. So, basically, we're using the one second square pulse protocol of the Allen Institute. And what you do is you inject several pulses of depolarizing stimulus and also several pulses of hyperpolarizing stimulus that doesn't appear here.
If you're familiar with patch clamp recordings, or intracellular recordings, you know that you could do this. So in this case, this cell is a very typical, fast spiking cell. What you see here is just the raw traces. And here, for each one of the stimulus intensity, the spike rasters. These cells, when you stop the stimulus, the cell stops spiking. So there is no persistent activity.
But we found other cells that they look very interesting. For example, what we call a rebound persistent activity, which are cells that they don't have persistent activity when you inject the polarizing pulses here in green. But when you inject hyperpolarizing pulses, the cell have rebound spikes, here.
And here, you see this big H current, if you're familiar with these traces, that is very prominent in monkeys as much as in humans. And the recent paper by the Allen showing that these currents actually are very prominent in humans.
And here, you have a cell that you stimulate with, actually, excitatory pulses and you see persistent activity after the cessation of this stimulus. Here, for some of the pulses, not for all of them. There is a sweet spot that you can at least [INAUDIBLE] persistent activity.
Doing this, we found that, actually, persistent activity is in about one-- four of 164 neurons that we patches in five animals. So, basically, the interesting thing here-- or the thing to do here is try to classify those neurons and see if there is anything special with the morphology of the neurons.
We injecting them with biosetin. And then we actually tried to look at the reconstruction. This reconstruction is not quite ready. We still have to reconstruct the axon. And we're trying to classify those neurons that they show persistent activity. So the answer to the two hypotheses is, well, it is complex.
So in vitro, we find neurons that open cessation of the stimulation. We have persistent activity. And this is blocking synaptic transmission. So it is a complex question. Now, we publish a paper not long ago with Diego in which we look at different brain areas.
In some areas, you have persistent activity. In some areas, you don't show much of the persistent activity during working memory task. And there is this paper by Boston-- here in Boston by Jennifer Luebke where she is looking at, actually, characteristics of pyramidal neurons in the neocortex.
This is in the mouse-- compared V1 and prefrontal cortex. And this is in the human-- compared V1 and prefrontal cortex. And what we-- I'm sorry, in the monkey. And what we see is that this pyramidal cells grows-- like, almost is a giant relative to the V1 neuron in the lateral prefrontal cortex.
So the morphology of the different neuron types-- it seems to be, that has evolved quite a bit in primates. And, actually, we don't know the whole truth about this. And we are trying to do more research to understand cell types and how they relate to these functions that we are looking in completion.
Now, what other features of cortical circuits may actually make persistent activity to arise? We're looking into a model-- Chao-Ching Huang, that is basically based in neuron types. So neuron types in the cortex are very diverse. And they're connected into a certain way that, luckily for us, there is a lot of tools out there that we're trying to explore this question.
Chao-Ching's model, actually, that it was taken from his work with Patricia Goldman-Rakic, too-- so proposes that you have here, in gray, the pyramidal neurons with the optical dendrites. And you have three type of interneurons.
The parvalbumin interneuron, the calretinin interneurons, and the calbindin interneurons-- these are calcium-binding proteins that, if you actually tag the neurons with certain antibodies for these proteins, you can segregate the neurons into almost three different exclusive groups. They're not completely exclusive. But more than 90% of the cells you can classify according to this.
So the whole idea here is that the parvalbumin-- the pyramidal neuron excised the PV interneuron and the PV interneuron send feedback inhibitory signals to the same neuron. And basically, to the neurons here on this side, more strongly to neurons encoding different features as these specifically neurons that is activated.
And now, this synapse is into the calretinin interneuron. Now, the calretinin interneuron almost exclusively synapse into the calbindin interneuron. And the calbindin interneuron controls the flow of excitation through the optical dendrites.
This is what we call in the calretinin interneuron-- so inhibition of inhibition because it-- basically, it's an interneuron phone that facilitates pyramidal cell excitation. And you have, now, the PV that inhibit the pyramidal cells.
So the whole idea that he was proposing was that you have an increase-- in association cortices, you have an increase in calretinin interneurons. And, therefore, use half the inhibition of this neuron that actually shuts down the optical dendrites of the pyramidal cells. And you can have a flow of inputs. And that may actually favor recording excitation. It's a kind of a simple idea. So that was the idea.
Now, what can you do as a system neurophysiologist? You can actually look at the width of your action potential. Luckily for us, actually, the narrow spiking cells-- most of them are parvalbumin interneurons.
The broad spiking cells, I don't think that you can tell much about whether they are pyramidal, calretinin, or calbindin, rather than because they're more pyramidal cells than anything else, the probability to be a pyramidal cell if your broad spiking is larger. But I don't think that you can tell much about it.
So what we did was, basically, we sorted our spikes in these three different areas. We classified the action potentials. Here, you have a narrow spike and a broad spiking cell, very clearly. And we look at the proportion in the different areas. And here what we found.
What we found is that-- here is the histogram with the spike width. So what you see here in blue is the broad spiking cell and in red the narrow spiking cell. And for an MT to lateral prefrontal cortex, you see that the proportion of narrow to broad spiking grows smaller. So, basically, you have less narrow spiking neurons in the prefrontal.
Now, you could say, well, you have less-- fewer PV neurons in the prefrontal. Well, that's not fair because when you put on extracellular electrodes, you have sampling bias. You have a lot of things happening in electrophysiology that we know that it's just-- it's a stretch.
So we decided that we're going to do immunohistochemistry and we embarked into this journey with some of my colleagues. This is just to show what I've shown before, that there is an increasing putative PV cells. But we don't know anything about these neurons.
So what we did was, again, some of these animals were going into terminal experiments. So, actually, we made slices of the three different areas. And we actually-- here, you don't see that very well, but there is a [INAUDIBLE] stain here that defines layer 4. And here it is actually, the different areas-- MT, and MST, and lateral prefrontal cortex.
MT and MST is good because you can fit it in one single slice. So that actually good. And then we-- it took us some time to troubleshoot the antibodies. What you see here, the little dots, are the different neuronal types. So we have antibodies for the four main cell groups neurogranin for the excitatory cells here.
And, actually, maybe that slide you can't see very well, but, probably, the next slide is going to give a summary of that. What we did was basically count neurons-- a large number of neurons here-- the sample size in each one of the different areas, and a large number of samples in each one of the animals.
And we came up with the distribution of the number of PV, calretinin, neurogranin, and calbindin, and actually, more importantly, the ratio of PV to calretinin. So this ratio actually goes down as you go from MT to MST to lateral prefrontal cortex.
So, actually, we can say that the number of calretinin to parvalbumin-- or you can say in the other way parvalbumin to calretinin-- basically changes from MT to LPFC. Because calretinin increases in number, we deduct that there is more inhibition of inhibition in lateral prefrontal cortex, according to the number of cells.
And that may be a feature of cortical microcircuits that allows sustained activity to arise. So, again, that's basically the same review that we were talking about, that in these association areas, like here, for example, IT, here you find the prefrontal cortex, you have sustained activity.
But when you look in primary sensory areas-- here, in visual cortex and here, in somatosensory cortices-- actually, you don't see that often, the sustained activity.
Now, what is the link between persistent activity and working memory? In this case, we switched to a series of experiments. And the first question that we asked is persistent activity in population of neurons sufficient to encode working memory? So that's the criteria. It may not be necessary, but is it sufficient to encode working memory?
And we implanted [INAUDIBLE] microelectrode arrays in the lateral prefrontal cortex of two monkeys. And then we trained the monkey in a very simple task. So this is the monkey brain, as many of you recognize. This is the lateral prefrontal cortex. This is the array.
And this is interoperatory images that we corroborate the location of the array so that we could do a good mapping after all. The task is pretty simple. It's an oculomotor delayed response task. The animal fixates. So a target appear at one of these 16 different locations.
So the dots is to-- the dotted line is to just signal the locations. In reality, there is a blank screen only with the target. Then, there is a delay period. And the animal, when the fixation point goes off, [INAUDIBLE] to the remember location. It's a typical working memory task that you actually-- you cease the activity during the delay period in one example neuron.
This is not surprising. I mean, this is very well known. And this is actually the receptive fill of these specific neuron for the 16 different locations. Here, these cells, for example, light the right lower quadrant. So has the memory filling the right lower quadrant.
Now, what we did was to use a machine learning procedure in which we actually tell the best cells that we have with the best selectivity. And actually, we start pairing these cells with every other single cell that we recorded simultaneously. And we call it the best of ensemble method. And then, we start building ensembles of two neurons, for example.
And then we choose the best ensemble of two neurons. And then with that best ensemble of two neurons, we start pairing with any other neuron in the sample. So then we take the best ensemble of three neurons. I understand-- so let me make clear here, this is not really the optimal ensemble.
Because exploring this space is going to take us four years with the computer power that we have. So it was-- so what we are doing is pairing the best two with the other-- and finding the best trio that includes the best pair, and the best quartet that includes the best trio.
By doing that, we can reach a performance of close to-- in this case, to 80% with about 15 or 20 neurons-- the coding performance is about with 15 or 20 neurons. So our conclusion here is that, yes, the activity seems to be sufficient to encode working memory in these lateral prefrontal cortex neurons. Now--
AUDIENCE: But only when you take the population, not individual cells. Is that correct?
DR. JULIO MARTINEZ-TRUJILLO: Population. Yeah. Individual cells will make it--
AUDIENCE: Individual cells are not good enough.
DR. JULIO MARTINEZ-TRUJILLO: No. Not good enough. So when you take the population and you put into the machine learning into-- well, we use [INAUDIBLE] with logistic regression super vector machine, linear, nonlinear kernels. Nonlinear kernels was problematic because we're overfeeding the data. So we decided to stick to the linear kernels for now. So that's what we are doing.
But, yes, it's just the population. So there is a lot in this paper correlations and a lot of things that we look into it we can talk about that later, if you like to. I wanted to go to this exciting stuff about the monkeys playing video games. So, basically, do lateral prefrontal neurons encode working memory more realistic situation, regardless of eye movements and distractors?
So our experiments are usually performed in a very reduced environment. That's what-- scientists, we are reductionist by nature. Of course, we have to isolate the variables. If not, we can't control for them.
But, at the same time, if you take a system that has 150 variables and you reduce it to five, maybe the responses or the-- when you interrogate the system, what you're getting is not exactly the same thing as when the 150 are present.
So we develop-- looking into the video game development, we develop this tool box that basically communicate on Real Engine 3 with our software. And we develop, also, algorithm to detect eye movements and to measure eye movement during these virtual environments.
Unfortunately, you decay actually on Real 3's death. So you have to switch to Real 4. So I could tell you to download this, but it's kind of useless now because we are trying to upgrade this. And this is the monkey. So this is the monkey here, with the joystick. And the monkey's navigating to pursue, actually, this blue color.
Someone is telling me, so do the animals really think that they're immersive, and they are really in a virtual environment? So I'm going to show you something now that is mostly anecdotal, but it may convince at least some of you. There is the monkey going. And he's chasing. He's doing just really well. So now the monkey go for the red stimulus, and go there.
And I believe that is happening in the next one. So, basically, what we find is that the animal-- oh, no, it's not happening in this movie. Maybe I got the wrong movie. So, basically, what we find is that when the animal makes a mistake-- oh, this is not happening. Excuse me. The monkey is cheating.
At some point, the animal-- if I keep playing the video, probably, you would see, at some point the animal goes into the left one. And he backs up the wall and actually try to go to the other one. That's what it is in the video. This video probably have to go for a long time.
But I do believe that the animals are navigating the environment. They avoid obstacles. They go around them. So that's the evidence that I have, so far. Now, we created this virtual arena for doing a virtual working memory task. And in the virtual arena, it's just motivated by circular mazes. So we have the animal in there, in this podium. And you have these nine different locations.
The animal is basically navigating with the joystick. And the task is kind of simple. So you shine this stimulus somewhere. Then, you have a virtual gate that doesn't let the monkey to go. And then you let the monkey go. Let's say, now this goes off. He can go. That's a delay period. And now, he goes. And then he hit the location and he gets a reward.
The monkey does that again, and again, and again. And actually, he does pretty well this task. In this task, you can identify three different periods-- the q, that is about three seconds. The delay, that is about two seconds. And the response, that is about 10 seconds.
And during these different periods, you can ask two different questions, in this case, we ask. First, do lateral prefrontal neurons and populations encode visual memory working memory in virtual environments in the presence of eye movements and visual stimulation? Remember, we did control for-- we measured the eye movement. We didn't ask the animal to fixate.
And the second is more related with the causality between the activity and working memory, which is-- does changes in persistent activity produced by blockade of the NMDA receptors, using ketamine, actually produce working memory deficits in this animal?
There is a whole literature on ketamine, and actually working memory and ketamine and depression. But what we know about ketamine it's an MNDA receptor antagonist. So here, just to show you the little clinical trial that we run, here. So basically, we have the pre-injection trials, in which we measure the performance of the animal in all these periods.
Then, we have the injection of ketamine. We had to train the animal for doing this injection so it doesn't get a lot of-- very stressed out. And the animal keeps working. And you have the early post-injection and the late post-injection period. And you can do first at the performance of the animal.
Well, first of all, we implanted two microelectrode arrays in the dorsal and ventral lateral prefrontal cortex. And, actually, we use these designs that it is published in a paper by Blonde, et al. 2018, where you use these caps that improve a lot our incidence of troubles with the animal.
But just to recap-- so there is a ventral and a dorsal array in the lateral prefrontal cortex in the two animals. Those are [INAUDIBLE] array with 96 active channels, each one. So you could record quite a bit of data.
Now, what about the performance? This is the performance of the animal. So the number of correct trials-- correct trial is where an animal hit the location in a certain time window. This is the number of correct trials with saline injections here, in gray. So we repeated the injection because the injection could be an uncontrolled variable.
So that might just have been affecting the performance of the animal. And here with ketamine. Actually, I just want to go back to this. I apologize for that. And I just want to say that these were the doses of ketamine that we use-- 0.25 and 0.4 milligrams per kilo. We titrated the animals. We titrated the performance of the animal.
If the animal start getting nystagmus, we got to-- that dose is not going to work. So we have to go to the lower dose of ketamine. So this is the performance with ketamine. The animal, actually-- the number of correct trials went down. And then it recovers, actually during the post-- the wash out period, in about close to an hour.
And this is the reaction times to reach the target. They increase with ketamine. And they recover during the period relative to the saline control. So, in this case, we do have a deficit in working memory-- performing a working memory task in the animal. Now, how does it look like? This is the trajectories of the animal from the podium, all the way down to the target.
And they look pretty regular there. When you look at, under ketamine, all these black trajectories-- the animal's just like roaming around. And it kind of forget where the whole thing was. And sometimes it hits it, sometimes it doesn't.
And then, during the recovery period, it seems to normalize a little bit. So somewhere, the animal is just wandering around in the maze and he doesn't find the location of that.
So the whole idea that we proposed was that ketamine, basically, acts at the level of this synapse here, mainly in parvalbumin interneurons that contain this type of receptor. And where low doses of ketamine may actually impact mainly this synapse. And what happened is that the lateral inhibition of neurons and the inhibition of the same pyramidal neurons doesn't happen.
And therefore, the tuning curves actually flatten out with ketamine. So the tuning curve of the population is what gives you the tuning curve for the memory signal, of course. So this is an example neuron, in which we have, here, in different colors, the firing rate of the neurons during the delay period.
What you have here is a plane that fits these firing rates and actually gives you, for example, in this case, this neuron was more selected for the target that was proximal and to the left when the animal comes here. So and these are the different-- super smooth-- actually, it's by density functions.
Now, when you give ketamine, what you see is that the response to a lot of these non-preferred directions start going up. And this thing becomes yellow. So basically, what you have is just a flat surface. And now, in this case, after the ketamine, you have the recovery where the tuning looks a lot like the tuning during the pre-ketamine period.
So the conclusion here is that during ketamine, relative to saline, a lot of the neurons-- what you see here in this pie graph is the pre-injection selectivity. Here, the percentage of neurons known selective. And here, the percentage of neurons selective for the different periods of the task.
What you see is that the percentage of selective neurons for the different period of the task decreases a lot during the early post-injection period relative to the pre and then recovers during the post-injection. That doesn't happen with saline. With saline, you see that these proportions remain very similar.
So, basically, we can conclude that ketamine actually decreases the tuning of LPFC neurons. Now, you can put all this data that we're recording simultaneously into a classifier-- a linear classifier. And you're trying to classify the remembered location in the different task periods.
So what you see in green and blue is the performance of the classifiers during the different period-- cue, delay, and response-- actually during the pre and the late post-injection. And during the early post-injection, injection the classifier actually drops in performance. We're investigating why these drops happen.
I suspect that it's a loss of the signal, rather than noise correlation, so spike on correlations. But we have to investigate that more. Now, interestingly, if you do that with saline, the classifier doesn't drop in performance during the early post injection period. So we have evidence that the single neuron at the population level that this is happening.
What about eye movements? I know that many of you are dying to ask me, but you have to gain some constraint. So here is the thing-- their mind's looking everywhere. So it doesn't matter if the animal is navigating to there. He's not going like-- navigating to that position. He's just navigating there and he's looking everywhere. He's just taking his time to look everywhere.
So I don't think that we can predict from the eye movements, actually, from the eye fixations where the animal is going to go. This is just-- I want to back up this statement. We took, actually, the eye positions, and we put it into a linear classifier-- the eye position-- to decode the location of the cue, and where the animal was going during the delay period.
And the classifier is almost like chance. So, basically, the way that you do this, you take the eye movement and you try to classify where the animal is going from these nine different location from the pattern of eye fixations. So the classifier is almost like chance. If you look at the neural data, the classifier is almost-- is not a chance.
We compensate for the number of features in these classifiers-- number of neurons and number of features that we enter in the classifier. So, basically, I don't think that the eye movement is a very reliable estimate of where the animal is going if we were thinking that.
That's bad news because we use a movement a lot in many of our tasks, including my lab. So sorry to be the bearer of the bad news. That doesn't mean that-- well, anyway, not going to talk in the module. Probably, you're going to ask questions.
So a summary-- LPFC neurons encode working memory in naturalistic conditions in the presence of eye movement and distracting information. So ketamine reduces performance in the working memory tasks and the spatial tuning is lost in individual neurons. And the ketamine decreases the accuracy of population decoding of target locations.
I think that I started about 4:15. So I still have 15 minutes. And I want to talk to something that may be interest to some people in the audience of this coding of long-term memory by single neurons and ensembles in the hippocampus.
[VIDEO PLAYBACK]
- During life as a brain surgeon, it has been necessary to operate on a good many men and women-- good many hundreds-- and expose the brain, under local anesthesia, with the patient conscious. In those operations, it is a useful, practical, procedure to stimulate the cortex electrically.
These are not experiments. In that process, we have stumbled, quite accidentally, on the fact that there is, recorded in the nerve cells of the human brain, a complete record of the stream of consciousness, all those things in which a man was aware in any moment of time.
[END PLAYBACK]
DR. JULIO MARTINEZ-TRUJILLO: So I'm going to stop here because, probably, many of you have seen this video. That's Wilder Penfield. So those were the times of the [INAUDIBLE]. He said, they were not experiments. They were just findings-- casual findings.
So, after that, Brenda Milner-- actually, this is a patient-- HM-- Henry Molaison-- that has a bilateral resection of the temporal lobe. And many of you knows the seminal studies on Brenda Miller, which is one of the pride of Canada. So is Penfield.
So, yeah. I have to stop every time to say how incredible she is. I just can't get over it. So she's just an amazing person and researcher. So she published these paper with-- that it becomes an icon of people doing memory research, just basically.
So the whole idea was that it was lost in the ability to form memories, especially declarative memories, or associative memories in this patient. And people associated the hippocampus with the memory field, obviously. But in the rodents, a little bit of a different picture arises.
So John O'Keefe received the Nobel Prize for describing place cells in the rodent hippocampus in 1971. So that was what started the research. And basically, when the rodent navigates through a maze, there are place that responds on a certain position in the maze. So this is very critical.
There is now a split in the field-- kind of a split, I find, between people that do spatial navigation and people that do memory research. And this kind of touches a little bit-- the humans, and non-human primates, and the rodents literature-- and the rodent researchers, so as many of you know.
So one of the questions that we wanted to answer, by going into the hippocampus and look at the nature of the representations in the hippocampus of the memory representations, was-- how does mnemonic coding of space and visual feature interact in primate hippocampus neurons during virtual navigation tasks?
Of course, visual navigation became available at this time because we have, now, the virtual environments. The monkey can use a joystick.
And it was also a development of brain navigation techniques that some of the companies that were based in Montreal at that time wrote research-- work with us and we could actually develop a very good targeting of the hippocampus, even with subfield resolution that we are doing at this point.
And, at this point, we aim to target the hippocampus in a virtual environment using different tasks-- navigation tasks and associated memory. This is the X maze, which is inspired by rodent research. This is a mountain. And this is a tree. Don't ask me why they're there. I'm still asking this question to Rob, my grad student that put them there.
Why do you want them there? Oh no, I like them. Apparently the monkey likes the trees more than the mountain. That's the only thing that I can say about that. But what I'm going to talk about is not about allocentric cues now. So you can think about allocentric cue, but what I'm going to talk about is about a little bit of a different paradigm.
So here is the X maze. In one task, we train the monkey just to hunt for an object that is going to be rewarded. And you position the object at any different position in the maze. There is no trial structure here. The monkey navigates. Hit the object. Get the reward. Keep going, finding, keep going. So that's what the monkey does.
In the second task, what we call it the associative memory task, we anchor the targets here, to the arms of the maze. So the targets could be these colored dots. And the walls of the maze would be either wood or steel. And dependent on the wood or steel, the animal has to figure out this color scale.
And the animal has to navigate to the item that give the animal the highest reward. So if it is wood and you show green and orange, animal has to go to the orange. But if is actually steel, and you show the animal green and orange, the animal has to go to the green one. So the animal has to reverse the context association.
They do that very quickly. This is an example trial of one animal. You see is not spiking. You got a spike there. So it amazes me how these hippocampus cells actually fire. It's just a little bit of a crazy thing. So on the cell, seems to be spiking-- there, as you see, this more random spiking or, I'm sorry, sparse spiking.
So this is the cell where the animal is just foraging the task. And in the second task, the animal is actually doing the associated memory task. So what you see is the eye movement. And you actually can recover the object that the animal is looking at.
So here the walls are steel and the animal has to go to the green target and bam, he gets a reward. Now the animal goes back. And this is amazing. Every time-- I mean, some cell is amazing. But I'm going to talk about that.
So here is the wall. The animal goes back. So that context is wood. The animal has to go to the orange and not to the green one. It's very clear, the task, right? It's just a context association task. And again is the cell firing after the animal receives the reward, the animal goes away and the cell keeps firing.
So how the cells firing the two different tasks-- we could hold the cell into two different sections. For example, here you have a cell that doesn't fire in the foraging task. So here is the distribution of the positions on the maze and here the spikes rasters for the different positions in the maze. In this case, the cell doesn't fire.
But when you go into the associative memory task here, in the arms of the maze where the objects are appearing, right, where the different objects appear-- so the cell fires. In this cell, they fire a little bit here in this task. But in the other task, the cell fires again here, in the arms of the maze, and a little bit here in the corridor.
And this cells seem to fire in the corridor in this task. And in here, it's just firing everywhere. So, basically, there is quite a bit of change. The same cell change firing, when you have the animal in the same context. The animal is navigating the same space, but now you have changed the task contingencies. And now this cell changes.
That reminds a little bit of how play fields actually change in the hippocampus literature when you change the context. So what you have here is the distribution of the different what we call place fields in this animal in the foraging tasks that is all around the maze. Here is the proportion of place plays fields for different areas of the maze that we have in different colors here.
So what you see is that is kind of homogeneous in the foraging task. Now, in the associative memory task, what you have is, in the arms of the maze, you have a sudden appearance of what we call place fields in these areas of the maze. Are those really place fields? I'm not sure.
So we sent this paper and some-- people from the rodent groups actually made us to do a lot of spatial information content and a lot of things. At some point, I was accused of being actually like a dirty-- doing kind of sloppy rodent hippocampus research.
So I want to talk about, that is traumatic. So now we have refined our analysis. And we have done a lot of the analysis that they do in the rodent, which is true. They have developed a lot of this analysis very thoroughly. This is a field that has evolved a lot. But the results are the same, whether you deal with spatial information content or we do it with actually looking at rates.
So what mediates the changes in single neuron in factual information contents across that? I just want to take you to this part of the task where, actually, the objects appear right in the arms of the maze. So this is exactly where you see these resurgents or this appearance of place fields right in this part of the maze.
Now, we have three variables. One is the context. The other one is the objects that we have in every single trial. So what you could do is basically divide the maze into three-- into different periods that correspond to different spatial areas. So post-reward or trial, pre-context, context appearance, here in the corridor, object appearance in this case, and object approach. And the trial ends, and the animal goes again and loops through that.
So, now, these are the responses of several neurons to different of these during the associated memory tasks, specifically. What you see is that actually, for example, in the object approach or trial end, this neuron is firing a lot of spikes. Here, in the x-axis, is the firing rate of the neurons in different trials.
You see that there is a lot of spikes fired here, but not in the other areas of the maze. But these neurons, you have the opposite thing. That is happen, actually, when the animal is the post-reward trial. You can interpret that as the animal, here, is perceiving the stimuli. Here, the animal, maybe, is rehearsing what kind of object pair association was shown in the maze.
So we did a linear regression in every single neuron. And what we see is actually that the Beta coefficients for these specific areas of the maze-- object approach, and in this case, post-reward-- actually are increased. Let me explain this, maybe, a little bit better.
This is what we call trial N. That's when the animals start a trial. They see the different-- the association and the targets and then they get the reward. And this is the trial N plus one. Where we are regressing here is the context object association of the previous trial. So this is actually retrospective coding.
So there is no object shown in this post-reward. The object disappears. So our conclusion is that maybe the animal is rehearsing what he saw in the previous trial, something that you find in the hippocampus literature.
So what we believe that is happening here is that the neuron encoded behaviorally-relevant stiumli of the associated memory tasking the arm and branches of the maze, and they encode it in two ways-- perceptual and mneumonic.
I'm going to go, maybe, quickly through the next slides. So what we tried to do is to put all what I have told you about the spatial decoding in single neurons and featured decoding or context object decoding in single neurons, and now we bring it into classifiers.
So, basically, we produce pseudo populations and we tried to classify one, spatial position of the animal in the animal in the maze, two, what kind of trial episode it was-- was wood, plus green and red on the left side-- green on the right and red on the left side. This is just a classified performance.
Maybe you should focus here and there in the middle row. This is a classifier performance relative to chance performance that we use. We do it by shuffling the trials. And what you can do is that you can classify the position of the animal in these different areas of the maze using the population firing rate with a level that is higher than predicted by chance.
And you can do that in both tasks. But this is an allocentric coding, basically. We use an allocentric reference frame. Now, if we convert this allocentric reference frame into a direction dependent frame-- in this case, a trial that goes, for example, left up, here, to right down, and a trial that goes from left-- right down to up is considering the same thing because you kind of linearize the maze.
And you consider that's like an egocentric frame in which the animal is navigating from the post-trial, from the reward, toward the next goal. Your performance increasing substantially. So that's actually in both cases. These are the confusion matrices. That maybe we should focus on this for now.
Now, the most interesting thing here is that the performance-- the classifier doesn't generalize. So when you train the classifier during associative memory and you try to test it during navigation foraging tasks, it doesn't generalize. The neurons change the firing rate. Something happened that the classifier doesn't generalize.
So our interpretation is that the spatial code changes from one task to another. So this code is not stable. It's task specific. Or maybe it's stable for a task, but not across tasks.
Now, again, what we wanted to do-- and this is the last piece of information that I have-- was to decode trial episodes from the population activity. I don't want to talk about episodic memory because episodic memory is very hard to define, but about trial episodes. Can I decode that it was this context, and there were two different, actually, green and orange, in this case?
The answer is, actually, yes. So we produce here the predicted label for the different combinations that we have of context and object. Here, for example, wood. Here, steel context. Here, the different objects. And here, we have the real label. This is a confusion matrix.
And we have a very nice diagonal of the confusion matrix, where we can actually decode both chance, the trial episode that we show to the animal. And in this case, we decode that, also, prospectively, which is during the time period that the animal is approaching the target, we integrate fire rate over all this period.
Or we can encode it retrospectively, when the actual stimuli went off and the animal is going back to do the next trial-- during that period. Actually, this is prospective and the other one is-- so which is one-- yeah, this is the memory-- retrospective and perceptual.
And you increase, actually, your decoding accuracy, in this case, if you join all these-- the two periods. Here, we have to compensate for the number of features of the classifiers, if you can imagine, because here you have two time periods time number of neurons.
So but the conclusion here is that we could actually decode the trial type from the neuron. So to conclude, I believe that single neurons and ensembles in primate hippocampus is called a spatial position. This code is very similar to the one that they have seen in rodents. However, this spatial code is not specific-- it's specific to a particular task.
So the cells within the same session change the firing rate when you change the contingencies. The neuron ensembles in primate hippocampus encode information about trial episodes prospectively and retrospectively. And it could be-- this could provide building blocks for episodic memories.
Now, I have to say that a person that guided us, really, in our soul search in this hippocampus research was Howard Eichenbaum. I was really heartbroken when I heard the news. I think all what I can say about him that it was the most kind person that I have known. And he predicted this.
So the role of the hippocampal in the navigation of memory-- he thought about the hippocampus as this multidimensional device that actually encodes time in many dimensions. He was really an incredible mentor to my student and to me, too, actually.
Sorry. I should actually shut down the volume, here. So these are the people that did actually the real job. So we have a happy lab. We do a lot of science. And then we tried to party hard also, sometimes. I have to say that here is Diego. Diego you don't recognize yourself here, probably.
That was in Montreal with our old set up. We have the old [INAUDIBLE] at that time. And here are the other guys. This is a neurosurgeon that has collaborated with us.
And here, a picture of different guys in the lab. Rob Gulli, they lead a lot of the hippocampus research, and Megan and Marion, actually, they paired together to do a lot of the ketamine trials. So without these guys, we haven't been able-- we would have not been able to do this. So thank you very much to everyone and to our funding agencies.
[APPLAUSE]
have an interactive transcript feature enabled, which appears below the video when playing. Viewers can search for keywords in the video or click on any word in the transcript to jump to that point in the video. When searching, a dark bar with white vertical lines appears below the video frame. Each white line is an occurrence of the searched term and can be clicked on to jump to that spot in the video.
