Electrophysiological and optogenetic characterization of feature attention and working memory across the primate cortex
January 17, 2023
December 7, 2022
All Captioned Videos CBMM Research
Visual attention and working memory are two different cognitive functions. However, because of their close relationship and interactions, it is often claimed that they share the same underlying neuronal mechanisms. Here, I will first describe results from experiments using multi-area neuronal recordings and large-scale optogenetic inactivation of the lateral prefrontal cortex in macaque monkeys to characterize the neuronal mechanisms of feature attention and working memory across multiple visual processing stages of the cortex. I will present evidence that feature attention and working memory have dissociable neuronal substrates. Lastly, as part of an effort to characterize the roles of specific cortical layers in feature attention and working memory, I will present the discovery of the first ubiquitous laminar motif of neuronal activity that is preserved across the cortex, characterized by opposite gradients of local field potential power in the alpha-beta and gamma frequency bands. Based on this finding, I will propose a Spectrolaminar Framework for the electrophysiological study of the cortex, and I will present FLIP, a fully-automated Frequency-based Layer Identification Procedure we developed.
ROBERT DESIMONE: It's a pleasure to introduce my own post-doc here at the CBMM series. Diego came to MIT from McGill where he had worked with Julio Martinez-Trujillo looking at the processing of motion and its modulation by attention and working memory.
And when he came here, since we had had experience with optogenetics, he decided he would take on what has really been a long standing issue in the field, which is are attention and working memory basically the same thing. When I lecture about this in my graduate course, I say, eh, pretty much is the same thing. But he decided he would try to use optogenetics to see if he could dissect the mechanisms for attention and working memory.
And that required really a major effort in technology development to try to find a way to modulate cells optogeneticcally over a really wide area because unlike in a mouse where the cortical areas are very small, or a subcortical structure, or even a retinotopic structure where you can target one part of the visual field, in prefrontal cortex, you have to target really large areas to do something meaningful. And so he set out on quite a quest to develop the technology for widespread optogenetic modulation. So we're really happy to hear about him telling about his results today.
DIEGO MENDOZA-HALLIDAY: Thanks, Bob, for the introduction. And thank you for the organizers at CBMM, Chris, and Hector, and others that made this possible. So basically, what I want to do here is tell you about the main project that I came to deal with with Bob here that he was giving you a hint of. But also tell you about all these really exciting lines of research that derived from this project that were not even in the plans originally that led to really exciting collaborations with different labs around the department, and to findings and development of methods that I wasn't even planning to do here.
So I've really had a great time in this place, not only doing what I wanted to do thanks to Bob's support at all levels to really imagine an experiment and get it done no matter what, but also to be surrounded by such an amazing set of colleagues with such levels of expertise and domains that I didn't know about before very much. So let me start off by-- so basically, these are the different projects that came about but let me start with the main project here. And I want to start by talking about the brain as a filter. I think of the brain really as a filter at various levels.
So picture yourself somewhere like in Times Square sitting down somewhere in the staircase eating a lunch. So what you've done so far is actually you've chosen a place to sit to look at the scene. Right. And you have a dome view of the Times Square but you chose to do that so that your first filter was head direction and gaze kind of filter. You chose where the range of your eye gaze is going to be.
A second filter starts with cognitive signals that we call attention, the modulation, the selective modulation of sensory processing that allows you to preferentially process certain regions and certain features, in essence a subset of all the sensory inputs that overwhelm the eyes and all the senses. A third level of filtering happens when you finish your lunch, and you're walking away, and now you're carrying with you some interesting stimuli that not only caught your attention but that you find you want to take with you. And that basically is what working memory allows you to do. So basically, there's a subset of those attendant stimuli that you are able to now hold on to despite the visual stimulation being gone.
And there's a fourth level of filtering whereas if you go back to Times Square a year later, you can remember that you were actually taking a picture in that same place of your friend taking a picture of a high tower in Times Square. And you might not remember anything else. Right.
So there's the four levels of filtering there. But I want to focus on two of them, attention and working memory, which you are very strongly related and yet I want to claim are far from being the same thing. And so let me start by telling you how strongly they are related first. And with that, I'm going to introduce you to feature attention, which is the kind of attention I study where you focus on nonspatial visual features like colors, directions of motion, shapes, et cetera.
So picture yourself at a society for neuroscience and some other conference where there are an overwhelming number of people and posters. And you're looking for that weird yellow poster that your friend says they made. So you're looking around. And you know it's yellow so you're using that-- you're using that feature to search around an overwhelming number of stimuli. And so what you do is you can actually preferentially boost the processing of everything that is yellow. And that will allow you to find that poster more easily.
And how this happens we believe is that you build an attentional template of that color yellow in working memory, which is what we call the process of attentional selection. And then you use that attentional template as a signal to modulate the activity of neurons and visual cortex in a selective way, making neurons that prefer that feature talk louder than the others, in essence.
And we believe that this mechanism of modulation might be subserved by feedback signals that may originate or may have a source in the posterior end of the lateral prefrontal cortex, called LPFC-p. And so many questions emerge from that. There's an increasing trend to-- I don't know if it's increasing because, apparently, Bob remembers it from a long time ago but I hear it more and more. People really claiming that attention and working memory are really one and the same and they share the same underlying neuronal mechanisms or substrates.
So I want to ask, first of all, are the neural mechanisms underlying feature attention and working memory the same or are they dissociable at any level? Next, I asked, does the posterior subregion of LPFC, that LPFC-p, play a critical role in feature attentional selection and in modulation? And does it play a critical role in the maintenance of working memory representations? And derived from that, we can ask whether working memory plays a critical role in feature attentional selection and/or modulation. What are the relationships between the functions and the area?
And so to study all this, I've designed a feature attention-- a working memory guided feature attention task that is spatially global in which macaque monkeys, which are our animal model of study, pretty smart creatures, are shown a cue stimulus of random dots that are moving in one direction and are covering the entire screen. So this is a global stimulus. And after a delay, they're presented with two patterns now that are overlapping and moving in opposite directions.
And animals have been trained to pay attention to the surface that matches the direction of a cue and report changes in speed in small patches that occur anywhere on the screen in the cue to direction and ignore changes in the distracting direction. And so if we focus on the delay and test periods of a task, we can compare two kinds of conditions with opposite cue directions. And in those two conditions, both the delay period and the test period have identical sensory inputs.
But what they differ in is, during the delay period, monkeys are remembering this direction versus this direction. And therefore, delay period activity allows us to assess the ability of neurons to encode working memory representations of those directions. And then during the test period, the attention period, monkeys are selectively attending to the cue direction and ignoring the other one. And by varying that cue direction, differences between these two conditions can then be assigned to an attentional effect on the neuronal activity.
So I recorded from neurons in multiple areas, first of all bilaterally in LPFC-p but also in parietal cortex area LIP, and in early visual cortical area MT, and intermediate level or visual association cortical area MST, neurons all of which had a motion direction preference. And so here I'm showing you a few profiles of activity of individual neurons that are representative.
And what I found was that while some neurons showed differences in firing rates between those two conditions that I was telling you about, so opposite directions, cue directions, during the delay period, indicative of working memory coding, but no attentional effects. Others showed strong attentional effects but very little or no working memory coding. And in fact, those two types accounted for the majority of neurons in all of the areas recorded. Only a minority of neurons had what one would expect if one thinks that the substrates of working memory and attention are the same, which is that those two signals are co-occurring in the same neurons. Only a minority showed that.
And you can see here basically coloring the neurons, the attention only working memory only neurons in red and blue and working memory and attention neurons in green. And you can see a clear dissociation, great dissociation between those two types of exclusive neurons. And that dissociation was particularly striking in LPFC-p, again, an area that has been thought of being a source for featuring attentional modulatory signals.
So next, I decided to look at the causal role of this area in feature attention and separately the causal role of the area in working memory. And to do that, I developed an optogenetic method in macaques for large scale minimally invasive optogenetic inactivation of the LPFC-p by injecting a viral construct containing the [INAUDIBLE] over a very wide area. And for those who are not familiar with the scale in the macaque brain, the mouse brain basically is twice the size of this.
So basically, this would be one hemisphere of the mouse brain. So it's a very large region in rodent scales. But it's basically the scale of functional areas in macaques. And I did that bilaterally. And I was able to then use an external laser stimulation through a transparent artificial dura that I implanted, again bilaterally. And at the same time I was recording activity from parietal and visual cortical areas.
Importantly, I inactivated LPFC-p bilaterally during either the delay period or the test period, independently assessing its role in working memory, and attention, and feature attention. So optogenetics worked nicely. You can see examples here of inactivation during the test period or during the delay period. And you had a reduction in the activity, so control in gray, green for opto, you see a nice reduction in the firing rates for this neuron. And this is the average of the population, the same for delay period of optogenetic inactivation for an example neuron and the population. So inactivation actually worked in the prefrontal cortex.
What was more surprising though was that when I looked at the activity recorded simultaneously from the other areas which are very distal from the frontal cortex, I found very robust reductions in firing rates that indicated some kind of more communication between lateral prefrontal cortex and the other areas. And that was true for all of the brain areas Recorded?
So that was promising. But I started first by looking at the effects now at the cognitive level, the effects on task performance. So during test period optogenetic activation, I saw a robust and drastic impairment in task performance and correct trials and an increase in reaction time with LBFC-p inactivation. And that was very promising.
Now obviously, that effect can be due to many non-specific effects. But luckily, I could do an interesting controlled experiment to kind of demonstrate the specificity of the effects whereby instead of bilateral inactivation and the unilateral inactivation of each of the hemispheres, and then I looked at the effects, the behavioral effects, in trials during which the change that the animal is perceiving happened in the contralateral hemifield to the inactivation or the ipsilateral hemifield, so contra versus ipsi. And we know that attention is lateralized in the prefrontal cortex. So one of each of the hemispheres subserves functions for contralateral attentional effects predominantly.
So what I found was that when I compared this ipsilateral condition versus contralateral condition, I found that optogenetic inactivation reduced performance for the ipsilateral condition as well but did mostly for the contralateral condition. So there's an effect of that laterally that is indicative of what we know about the lateralization of prefrontal cortex, demonstrating that this is likely an effect that is specific to vision and attention, not to any other non-specific effects.
So one question that emerges is whether the role of LPFC-p in attention is on the selection or the modulation components of attention. And one of the things that can give me some hints into that is the pattern of errors that the animals make during optogenetic inactivation. So we know that the target response trials, which are the correct trials, are going down with activation. But then we have different kinds of error trials that we can look into that would then compensate for that decrease in correct trials. And those error trials could be basically an early response or a late response, or no response basically, or a response to the distracting change.
And so these different errors indicate different aspects of impairment. For example, if we assume that LPFC-p inactivation is causing an impairment in attentional selection, basically selecting which of the directions is to be attended, you would expect that, in some trials, animals would make a mistake by selecting the wrong attentional template direction, and attending to it, and then responding to changes in the distracting pattern. So decreases in target responses would be compensated by increases in distracted responses.
Contrary to that, if the effect is on attentional modulation, you expect that the animals will still be able to select the right target for attention. But the modulatory mechanisms that enable perception to be boosted for that direction would fail. And therefore, while attending to the right stimulus, they would more often fail to see the change and end up having a no response, a nonresponse, which is this kind of pattern.
And what I observed in the data is, for both monkeys, the decreases in target response trials were actually accounted for by increases in the nontrials. And there was no significant difference in the early or distractor trial incidence, demonstrating or strongly indicating that the impairment due to LPFC-p inactivation is really on attentional modulation mechanisms, not on selection mechanisms. And importantly, those selection mechanisms are the ones that are most tied to working memory so they can start putting the puzzle together.
Importantly, when I looked at the effects of optogenetic delay period activation on performance, I fail to find any significant impairment on performance and reaction times as I had found with test period inactivation, showing an interesting dissociation at the behavioral level between inactivation during working memory and during attention.
So I wanted to look at the underlying neuronal effects of attention or working memory during inactivation. So I can measure the strength of attention in controlled trials and opto trials separately. And I'm showing you here two example neurons from LPFC-p and MST that show a reduction in the strength of, sorry, attentional effects within activation.
Here this effect is kind of reduced here. And for the population, I can basically do a normalized firing rate and show that, for LPFC-p and MST, that difference between preferred and [INAUDIBLE] direction shrinks. It causes a reduction in the strength of attention based on a ROC analysis, R-O-C analysis. And I fail to observe that in the entire population in MT and LIP.
Now if I further summarize that with these bar graphs, you see a significant drop in the strength of attention in LPFC-p and MST. But if I separate these neurons now by attention only neurons and attention and working memory neurons like I had shown you before, I find a really interesting additional dissociation where the effects of optogenetics are really limited to the attention only neurons. And they're present actually also in LIP in this case, whereas the effects on working memory and attention neurons were absent in all areas.
Now if I turn to the effects of optogenetic inactivation during the delayed period on working memory signals, I find that in LPFC, inactivation failed to reduce the strength of working memory coding, again measured by ROC analysis, also in LIP. But interestingly, I saw some effects in MST so the underlying-- so the reduction in firing rates came with a reduction in the working memory auROC. However, that was the only area that showed effects, significant effects.
And when I separate by neuron type, I find that those effects in MST were present in working memory only neurons and not in working memory and attention neurons. So basically in summary, the intended neuronal effects of optogenetic inactivation on attentional working memory are very different with strong effects on attentional modulation impairment of attentional modulation on neurons with much shyer and more specific effects on working memory coding.
It can also look at the effects of delayed period of optogenetic inactivation on attentional effects. And I find that there are no effects despite strong effects of optogenetic inactivation during the delay period. So to put all that together, I'm showing you that the neuronal substrates of feature attention and working memory are dissociable at the level of neurons and brain areas. And furthermore, that LPFC-p serves plays a critical role in attentional modulation mechanisms but less so in attentional selection mechanisms or in working memory coding.
So let me switch gears a little bit to a little tangent from that project that I wasn't expecting. It was an interesting finding that I got that I think has some important implications. If I go back to understanding what feature attention is, there's this really interesting model called the feature similarity game model of attention that really offers a beautiful yet, I would claim, somewhat wrong account of how you can use attention to boost perception.
So basically, the model says that when you attend to a given feature, neurons are up modulated in a way that is proportional to the distance between the attended feature and their preferred feature. So neurons that prefer that yellow that is being attended get the higher boost. And as that feature gets further away in feature space, the modulation gets smaller and smaller. And that, if you think about it, gives you a nice account of how you can actually change the population responses to boost preferentially the perception of yellow.
So I looked at my neurons. And I found a slightly different story. So sure, I found many neurons for which the preferred direction, which I can assess during the cue period, led to an attentional up modulation of activity during the test period whereas attending to the antipreferred direction led to less activity. So that's what I will call the upright effect. However, I saw too many neurons like this one where attending to the preferred direction led to a decreased response with respect to attention to the antipreferred direction. And that's opposite to what the feature similarity gain model of attention predicts.
Now I thought maybe these are just the exception. But when I looked at the makeup of the population across different areas, I found that in areas MST and MT, there was a drastic fraction of neurons, a very surprising fraction of neurons, that followed this kind of profile, shown here in green basically, with a lower than 0.5 auROC value with respect to the neurons that had an upright effect.
Interestingly though, as you move downstream to parietal and prefrontal areas, then the majority of neurons then have these aligned attentional effects that follow the feature similarity gain principle. And you can see that in the percentage of neurons with inverted preference being close to half in MT and MST and very, very few in prefrontal cortex. So basically, this data challenges-- sorry.
This data challenges the notion of the feature similarity gain model, claiming that the model does a fair job at accounting for the responses observed in prefrontal and parietal areas but fails to account for the diversity that you see in areas in early and visual association areas like MT and MST where some neurons show this effect and some actually flip their tuning during attention. So what mechanisms those imply is something that we need to develop and potentially model.
Now similarly, a parallel story comes with the persistent activity models of working memory, which really predict that the responses of neurons during the presentation of a visual stimulus are kind of carried over into a period in which the stimulus is removed and one has to remember it. And because of that carryover, it is expected that the code itself carries over from sensory inputs to working memory.
So I looked at that in my data. And what I found was that even though you see many neurons for which that selectivity is carried over into the delay period, I also saw many neurons that did the exact opposite for which the selectivity actually flipped and the antipreferred direction during the cue period became the preferred direction during the delay period. Just like with attention, you cannot ignore those guys. There are too many in the population, so almost half of the neurons in MST. And interestingly, much like with attention, as you move downstream into prefrontal and parietal areas, then you get less and less of them.
One question is are those inversions in feature preference, or they're really random changes, or is feature tuning just switching randomly? And I can test that by looking at-- by doing an interesting pictorial measure of a more precise preferred direction during the cue and delay periods and look at their difference. And so a zero difference means no change. And a 180 degree difference means an actual inversion in preference.
And so I looked at the distribution across all neurons of those differences. And what I see is that in an LIP and PFC, you see a unimodal distribution of those changes around zero, so showing basically little or no inversions, whereas in MST, I show basically two modes, one around zero, the ones that don't flip, and then one exactly at 180 degrees, showing that those changes in tuning really are inversions in feature tuning.
Another question that emerges is, basically, whether these effects are caused by interactions between neurons that are forced by having a global stimulus that drives a lot of neurons with different receptive fields in different parts and, through maybe normalization mechanisms, something is going on. So actually remembered I have some I had some data from long, long ago for my PhD with a similar task, a delayed match the sample task for motion direction where my stimuli, instead of being global, were local. And I had a sample presentation period like the cue period on a delay period of working memory.
So I analyzed those two in my data. And I found, again, in MST, neurons that had this preservation of preference and those that have had inversions in preference. And much like with my post-hoc data, the percentage of inverted preference during working memory was higher in MST neurons than in LPFC. So everything basically was replicated, which is a good sign. So again, it shows that the persistent activity models of working memory may account well for the response profiles in prefrontal and parietal areas but really failed to account for what is happening in visual association areas like MST.
So let me shift gears and tell you a little bit about another signal of interest that has been looked at less. So I've shown you so far what a feature attention does to spikes. But we also record often a signal called a local field potential that is really the aggregate of activity of many neurons. It is thought to represent synaptic inputs rather than spiking outputs. And it is oscillatory in nature. So we can actually look at how feature attention now modulates the amplitude of those oscillations in different frequencies.
So here I'm showing you basically the same task, the same types of analysis, but now with local field potentials in the frequency domain. So I'm looking at now a power as a function of frequency during the test period, the sustained attention period, for the two attentional conditions. So differences between them show attentional effects. And you can see in these examples attentional effects in the lower frequencies, lower frequencies, this one, and the higher frequencies, a large diversity of attentional modulatory effects.
To get a sense of what the population diversity looks like, what we do is-- and this work is really done by Aniekan Umoren sitting right there in the back. So thanks to him, we're getting all these really cool data. You can see that the activity, the strength of attentional effects in an area MT sites, which are now stacked, is predominantly in low level-- sorry-- in low frequencies. And as you move downstream-- maybe here I've organized the areas in a sensory to cognitive to motor trajectory.
As you go downstream in that processing stream, you see the attentional effects getting stronger and stronger in the higher frequencies and weaker in the lower frequencies. You can also see that in the percentage of sites or the proportion of sites that are modulated at each frequency, kind of increasing higher here. And as you go higher, they get higher here. And so we can actually summarize that by classical frequency bands.
And again, you see a really interesting second order effect of processing stage and frequency where the earlier stages have stronger effects in lower frequencies and weaker effects in the higher frequencies, showing kind of a downward slope. And that slope starts to flatten out and then become progressively positive now with MIP having the strongest effects in the high gamma and weakest effect in the lower frequencies. And that's shown here. So this is, again, an effect that is an interaction really of processing stage and frequency.
So this is the kind of complexity we're dealing with now with feature attention in the local field potentials, which we didn't deal with in the spikes. And we further showed-- we further saw that the population seemed to be separated into two different clusters. So one is these sites here that are primarily modulated in the lower frequencies-- you can see these here-- but not in the higher frequencies, whereas these sites here that were strongly modulating the higher frequencies were less modulated in the lower frequencies.
So that suggests there may be a dissociation between these two kinds of LSB sites. We actually did see that in across the population, the feature attentional effects showed a bimodal distribution across frequencies. So the strongest attentional effects are really either in the low frequencies or in the higher frequencies, or I guess the frequencies that are most modulated by attention are present mostly in the lower or the high frequencies. And we show that this profile is best explained by sites being modulated exclusively in the low frequencies or in the high frequencies rather than modulated in both with more than half of the neurons in every area but MST being modulated exclusively in one or another.
And what interesting again, kind of following what I had shown you before, those exclusively modulated sites in low and high frequencies showed an interesting trend across processing stages with those exclusive low frequency modulated sites being most prominent in lower stages here and becoming less, and less, and less prominent whereas the high frequency modulated sites, being less prominent in lower stages and becoming increasingly prominent in higher stages, so some kind of an effect again of processing stage.
And this leads to the possibility that these exclusively modulated channels in low and high frequencies might be separated atomically into layers, based on the observation that we've had, which I'll tell you more about in a second, that laminar compartments actually have differential opponents in the frequency domains with lower frequencies being more prominent in the deeper layers and higher frequency components being more prominent in the superficial layer. We are currently doing this analysis and seeing whether there's some anatomical distinction between those channels. And that will let us know more about the mechanisms that distinguish this and why they are this channels in the first place.
So let me shift gears to tell you about the ubiquitous spectrolaminar cortical motif we found. So one of the most fascinating features of the cortex is that, no matter where you go in the cortex, you see this beautiful laminar profile in the anatomy. And that was first described by Ramon y Cajal at the beginning of the 20th century but then it's been characterized and it's well known that no matter which cortical area you look at, there is this kind of six layer motif that varies across areas but is quite preserved. It's even preserved across different mammalian species with a cortex. So it really-- I mean you see this and you think, well, that's the key to the cortex. That's what we need to study. And yet we have miserably failed so far to fully find a truly ubiquitous pattern of neuronal activity that corresponds to those layers.
So we joined forces with the lab of Earl Miller at MIT with Andre Bastos, who is now at Vanderbilt, and a bunch of people that participated in this and analyzed laminar recordings from a total of 14 cortical areas spanning all sorts of functions. We started actually with two studies, the study of Andre Bastos at MIT and my study here with Bob analyzing that. And then the project kind of extended and included more and more people of which actually Alex Major is present, one of the people that participated heavily in this project.
So we analyzed the laminar recordings. So we're basically trying to record as perpendicular as we could from across all layers using laminar probes, V probes and S probes. And what we found was really interesting. When you look at the local field potential power spectrum, and you normalize it independently for each frequency, for each frequency bin, to look at which layer has the strongest power and which layer has the weakest power at that frequency, and you do that across frequency bins, you suddenly reveal a beautiful spectral laminar pattern that is preserved across different recordings that is characterized by a deep to superficial increasing gradient of gamma power and a superficial to deep increasing gradient of alpha beta power.
So we were interested. We saw this kind of crossover kind of pattern being common across recordings so we took the recordings of each of the areas for many probes and we aligned all the probe recordings to this kind of central point crossover point. And when we look at the average of those, of all the probes in each of the areas in each of the studies, we found how preserved the pattern is across all the areas.
So I'm showing you here four areas I recorded in my study, three areas that Andre Bastos recorded in his study. And you see how preserved it is across areas, across monkeys, and across studies. We actually used image similarity analysis to look at how similar these power maps are, relative power maps are. And what we found was also interesting.
The average image similarity of all within area comparisons was actually significantly higher than the comparison of between area image similarity comparisons, showing that despite the similarity, each area has some signature, some specific variance variability from other areas that makes it special and unique. We failed to find significant differences within versus between monkey comparisons, which is a good sign that it's preserved across individuals. And we saw a significant difference within versus between studies showing that some of the methodological differences might lead to differences but the image similarity was still quite high between studies.
We were not happy with the seven areas that we had studied, six areas that we had studied. And that's partly because there's a lot of variability in the lamination of a cortex, particularly in areas like the primary visual cortex that is heavily laminated with I think layer four subdividing into three compartments, and areas like the granular frontal cortex that lacks a layer four. So we really wanted to know whether this is a ubiquitous mechanism across the cortex.
So we included recordings from a whole eight additional areas, including D1, premotor cortex, somatosensory cortex, auditory areas. And in every single area we looked at of the entire 14, we found the same spectrolaminar pattern. So we're really after a ubiquitous property of the primate cortex, of the macaque cortex. But then the question was still there. Are these patterns of activity really layer dependent? Are they produced by specific layers or they have no link to specific anatomical layers?
So to answer this question, we ran some-- when I say we, that's really Alex Major sitting right in the front who did these fascinating electrolytic lesion experiments using even spaghetti that he purchased at the supermarket-- I don't have time to tell you what for. I'll leave you with the incognita. You have to read our paper now-- and then sent this over for histology at Vanderbilt to really look at where those physiological signatures, like the peak gamma, the peak beta, and the crossover were in the laminar anatomy.
So histology confirmed what we were already anticipating, which is that the peak gamma was mostly located-- this is LIP. This is PFC. And this is the average or the aggregate of all the areas that we looked at. And the data confirmed that the peak gamma was localized to layers two, three, whereas the peak alpha beta was localized primarily to layers five and six with the crossover between them being a nice marker for layer four, this in contrast to current source density, which if people know about studies of layers in the cortex, current source density has been the primary source of information about cortical layers.
We show that the relationship of the CSD sync, which is supposed to map the location of layer four, is actually really varied and not as robust as, for example, our crossover as a marker of layer four. Just to confirm this, I can show you the current source density maps in the same areas from the same recordings. And you can see how much diversity there is between these maps across monkeys, areas, and studies. This is also confirmed with image similarity analysis showing very low image similarity scores compared to what we had gotten with the relative power maps.
Alex Major showed nicely that the strict laminar pattern is robust to several recording and analysis properties, including the number of trials. So if you reduce to even five trials, he's still able to capture the spectrolaminar pattern. So you don't really need a huge number of trials to get a sense of where the pattern is and where the layers might be.
The pattern is also robust to sensory and cognitive states. And that was demonstrated nicely by comparing the spectrolaminar pattern taken from chunks of data from the cue presentation period of my task and Andre Bastos's task versus the chunks of data from the intertrial interval where monkeys are just sitting there looking at nothing, just doing very little, maybe thinking about their cage mates. I don't know. But that did not seem to change the observation showing really that this spectrolaminar motif is a basal property of the cortex that is present as a canvas over which all these phenomena happen. So we obviously want to look at the modulation, what small differences might occur with not only sensory inputs but cognition that might modulate the different aspects of the pattern despite the similarity.
The pattern is also robust to recording probe angle, as you can see in the comparison between the pattern in near particular recordings and high angle recordings. And that was not the case with the CSD. That kind of broke with high angle recordings.
And all this robustness led us to think that maybe this pattern is beyond science, that maybe a scientific observation, it can really help us design new methods for electrophysiology. And one of the things that we're thinking of developing is novel probe implantation methods.
First of all, we can use near real time probe mapping using the spectrolaminar pattern while the probe is being implanted in the cortex during manually controlled implantation. And so I'd like to call that the electrophysiologist's eye. And for any electrophysiologist, you know what I'm talking about. You can't see what you're recording. You can only hear it. And you can hear individual channels or the aggregate, but you can't really have this kind of laminar profile of where you are.
So now you can actually see that, as this example recording showed, where you're crossing a sulcus. You're coming in one cortical thickness here in one area, crossing a sulcus into another area. And obviously, they have inverted profiles. And you can see that inversion with the spectrolaminar pattern first being inverted as you enter. The pattern moves around as you keep entering. And then suddenly, the next pattern for the opposite lip of the sulcus emerges and shows the upright version of the pattern.
So you can really track a probe. And if we then do a closed loop microdrive control version of this, then we can think of designing fully automated probe implantation methods where you basically use an MRI coregistration and tell the microdrive I want to go right here. And the microdrive will basically detect these patterns and take you right there, not only for science but, hopefully, perhaps for medical implantation in patients.
So the other thing that comes out of that is that because there's a nice correspondence of the spectrolaminar pattern to anatomical layers, then we can use the pattern as a proxy to map the layer. So Nora Lee, a master's student in the lab, and I developed this fully automated algorithm that basically analyzes the data from scratch. So it takes your raw LFP signals and takes the power spectrum across layers, obtains the relative power maps, and then basically scans the data for ranges of channels where you see these opposite gradients of alpha, beta, and gamma that characterize the pattern. And with a measure of goodness of fit of these two gradients using regression analysis, we can then detect which of the ranges where the pattern is, find the peak of the gamma, the peak of the beta, the crossover, and map onto the corresponding or indicate the corresponding layers in a particular recording.
And this works really nicely. It even segregates our-- automatically segregates our probes into those that are very highly identifiable with high goodness of fit, those that are less identifiable, those that are really below an established threshold, and those that have these negative goodness of fit that indicate an inverted pattern, which occurs when you're recording when you're entering from white matter. So it also tells you additionally where you are in the cortex based on the orientation of a pattern.
So to summarize, this is the first demonstration of why ubiquitous or a truly ubiquitous laminar pattern of neural activity across the primate cortex. And it suggests a common layer and frequency-based mechanism for cortical computation across the cortex.
Now obviously, one interesting question is to test whether the pattern is also ubiquitous across mammalian species. So we were interested in looking at rodent recordings and human laminar recordings that we already got our hands on. And this will shed light on the evolutionary trajectory of these patterns and what they tell us about cognition, and human cognition particularly.
And last but not least, I'd like to tell you about a spectrolaminar framework for cortical electrophysiology that I proposed. And the framework is really trying to take advantage of the fact that if we now have a common laminar reference across different areas, then labs studying function in different areas and different functions can now use a common spectrolaminar reference space to map the results onto this space of layer and frequency.
So neurons, LFP channels, coherence between pairs would be represented not as aggregates across a given area but really subdivided into spectrolaminar compartments and analyzed that way. And so the hope is that with enough time, we can start building or understanding the similarities and differences between studies, between different areas, and with the hope of detecting what are the commonalities in the cortical computations despite the difference in diversity.
And the hope is that after we start building this framework, and when we start speaking the same language despite recording in different areas, we can start building towards a more generalized cortical theory that enables us to understand how variance of the spectrolaminar and cytoarchitectonic motifs in a given area enable it to specialize in a given function and how variations in that pattern across areas can give rise to the diversity of functions that are observed across the cortex.
I don't know if I have five minutes to tell you about the last story. I should probably wrap up there. So I'm going to skip the last really exciting story of optogenetic development that I did with Guoping Feng but I'd like to thank, first of all, Bob for all the support that he has provided and the support of many other PIs including Earl Miller, Guoping Feng, and Julio Martinez-Trujillo, my PhD supervisor, with whom I recorded that data that I showed you there, and all of the large team of people in each of the labs that really contributed greatly to everything that I showed you here. So this is really teamwork what I show you here. So I'll leave it at that. Thank you.
Associated Research Module: