Seminar paper

We will start I think we can start presenting ourselves while people still join

And then we can go on with the presentation So welcome everyone Thank you for joining this session My name is Sarah Romancy And I am a scientific application specialist at Iconos And I'm happy today to host our first journal club session

With our dear guest Doctor Felipe Sibis Pereira Felipe holds a double degree in chemical engineering from the Federal University of Santa Catarina and from the ESPCI here in Paris Then he obtained a master's degree from the

biomedical engineering in Paris program focusing on bioengineering applied to neuroscience And lastly in 2025 he completed his PhD in acoustic and neuroscience at the physics research Institute at the physics for medicine Paris Institute Who's also a founder of Iconos Currently he's a postdoctoral researcher

at the physics for medicine Paris Institute Focusing on functional ultrasound neuroimaging in free behaving rodents Something that of course he will also talk about today He's going to present his recent publication of January this year On Cell Reports titled a vascular code for

speed in the spatial navigation system Felipe will present And then we will have a Q&A session You can type your questions in the chat or in the Q&A And we will answer them afterwards I think we can start So, Felipe, the stage is yours

Hey, thank you Thank you Sarah Thanks for the presentation Do you hear me well Ok, yeah, so hi everyone Yes, so my name is Felipe And I'll present the paper of as for code for speed in the spatial navigation system I worked at as Sarah said I've done during my PhD at

the Institute physics for medicine in Paris in collaboration with icon use So one of the goals in this project was to Of course, to perform functional ultrasound imaging in freely moving rats As you can see by the drawing here And to explore cerebral blood volume patterns during spatial navigation

So I want to introduce here spatial navigation But for those that do not know what it is Spatial navigation is the idea that there are specific brain regions Especially the hippocampus and parietal hippocampus regions that have neurons that will encode spatial information related to navigation Like displacement

Displacement speed Direction, position to objects And so here basically we wanted to see if this spatial information that are encoded in neuronal activity were also encoded in the hemodynamic signals in the brain

So for the experiments themselves For the freely moving experiments We were so first we have a craniotomy made in each of the reds So we were using reds and would replace part of the skull by a polymer that is transparent to image during the o eight sessions And then the animals would be habituated to the

experimental setup to the experimenter And during an anesthetized session

We would select the plane of imaging We would select the iconous machine for

the ultrasound In total, we did 15 imaging planes across different animals And we imaged these planes chronically from a few weeks up to a month or two depending on the animal And all of them are sagittal because we wanted

to focus on capturing the pocampo The pocampo in parry pocampo regions And this is hardly like it's hard to do on a coronal's wise So our experiments were twelve minutes long and we would we have around in total at the end around

sixty imaging sessions in the in the project So this is what the experiments look like The animal would move around in the box with the probe fixed to its head And we would track a few parts of the animal's body And of course we would have the doppler video that we see you see on the right

And for those that are not familiar with the sagittal plane This is a sagittal plane where the anterior part of the brain is on the right and the posterior part on the left So in terms of data

from the tracking of the animal We can extract several different behavioral signals such as position Speed, velocity Direction and gravity in lots Lots, of different signals And from the doppler images We have here just to for as an example

We have some regions of interest just to show what the time series of these signals look like And for spatial navigation We are especially interested in the hippocampal regions here in light green on the left On the bottom left And also the parahippocampal that is well we have the middle internal cortex here in

magenta so the first thing and usually the most common thing to do when looking for patterns is looking for correlation right and in a in a correlation analysis what we want is to model why in a way that we can understand how much of the variation of x can explain the variation

in y and so this rate of explanation is given by beta here the coefficient beta n plus an error so I can model why as individual voxels so let's say in individual voxels why I

and for every voxel I'll have a model like this one so one beta per voxel since I'll have how much of a signal of interest x say the animal speed or the direction of the animal

How much of this signal of interest explains the variation in each single voxel yi So we can visualize y like this We have so for each voxel We can visualize like a animal speed So y, we have the time series of every voxel

So we can create a beta map like the one we are seeing here Here the values are very good because we are actually at lag equals zero Right We are y and x are aligned in time And we know that for functional ultrasound imaging Being in nemodynamic based modality

We have the nervos for coupling delay Right So this is so this is normal And if we put a little bit of lag between x and y We'll be able to understand for example how much of x explains y two seconds later Like with the lag of two seconds And by shifting this lag little by little

We'll be able to understand the evolution of the beta values as a function of this lag So from I don't know minus two seconds to ten seconds For example, this is what we did in the paper So if we look at this We'll be able to visualize the beta maps as a function of all lag And like we see here in the different

activation maps in the top In the top left This is not exactly the beta values here because actually it's the statistic associated to it And where high values are high correlations and negative values are negative correlations and here we have a threshold a statistical threshold that is done

So that's why we see so just some regions with higher low or low values So we see this surge of activation that peaks with within around two seconds of lag Mainly in the hippocampus in the hippocampus region

We have some of the important regions on the top right here that you if you want to see And then after two seconds it peaks and then goes back to the baseline Right So we'll see the videos because this we can see as a video Right You have a lag from minus ten to two to ten seconds We can see it as a video We'll see it in just a bit

But here in the bottom we have the average activation profiles for different regions in the in the hippocampus and parietal campus And yeah, so the thin lines are individual voxels from these regions And the thick lines are the

average signal on these regions Note also that on the bottom right We have they do not peak all at the same time Right So, and that's what we see here on the on the bottom right It is the same data But shown as a hip map

And we can appreciate the sequence of the lag to peak across these regions So it starts from parietal campus Then goes to hippocampus at around two point one seconds of lag So note that these activation profiles

can be perceived as estimates of the nerve aspergut plane function right for each voxel So we have one of these for each voxel And this will be important for field flights on the decoding part But just as an information here So if we do the same analysis for the angular head speed

So the rate of change of the animal's head direction while weaker than the correlation found for locomotion result We find similar correlations in the thalamus just under the hippocampus So we see that the hippocampus is not being activated here But we have the thalamus just

under the hippocampus and some parahippocampus regions to the side of the hippocampus as well And yeah, this is interesting because they receive inputs So the parahippocampus region receive inputs from the thalamus and they have

the both of these regions have had angular velocity cells or had direction cells on them So these two results are examples for a given animal But taking all our imaging planes for locomotion speed

We keep as the lag timings persisting as well across animals And what we see on the bottom right of the left panel here is also

interesting because the speed can it's correlated to other spatial parameters right So like acceleration or positive acceleration Negative acceleration And we show that while the regions might activate in a similar pattern

It's actually decoration to speed that is the most the strongest So we have the same thing for the angler head speed across animals and across imaging planes We have it's a little bit trickier with angler head speed because the

regions are smaller on the thalamus And while we change laterality of our imaging planes we might lose We might have some regions on one of the animals and not on the other So it's a little bit trickier to have this summary across animals here But we still have regions in the thalamus that are

coherent across animals and across imaging planes And we also show on the right that while angular speed can also be correlated to angular acceleration the absolute of angular acceleration And all that We also show that the angular speed is with angular speed that

we have the strongest correlation So we will look at the videos here across all animals and imaging planes And from the top left to the bottom right It's going from the central like sagittal and midline

The midline and to a more lateral to more and more lateral in the sagittal plane Yeah, and we'll see the videos We'll see that they all peak around two seconds and on the hippocampus and parahippocampus region And then the activation goes back to baseline

And here so if we freeze on the peak of the activation We see that across all animals and imaging planes We have this activation in the hippocampus and the

hippocampus have very low activation But we see the activation under the

hippocampus on the thalamic part or on the parahippocampal region anterior to the hippocampus So to further understand the relationship between the behavioral signals and the functional ultrasound data We wanted to see if we could decode this

behavioral and in particular continuous behavioral because we know we have the neurovascular coupling So we have a lag here and we don't Yeah, so we have some challenges on decoding continuous signal But then, okay

So we can, we just saw that we have actually estimates of the neurovascular coupling for each voxel Right And we can integrate them into the decoding index and we can then convolve this function of voxel

We can convolve the functional knapsack coupling estimates And this will actually realign this so the convolved signal will be realigned with the

behavioral signal for each voxel in the most not the most optimal way but in a way that the decoration analysis is telling us to So if we do this for animal speed

We then use quite a simple decoding pipeline actually to decode animal speed and you see here so here we have the weights for each voxel we have the weights of this voxel that the decoder

assigned during the during the train and here an example of the true animal speed in dark blue and the predicted animal speed in magenta using this decoder and only the functional ultrasound data So this is an example of one animal And we did this for all animals

And we had So actually we evaluate the goodness of the decoding by comparing our framework with a dummy regressor Right That would always predict the mean animal speed For example, but that would be like a fair guess for a dummy regressor And we get Yeah, so we get quite good results that are

way over 5 percent statistical significance here And we also did this for

Actually some spatial parameters that are if the animal is in the yellow region or in the blue region in the in the second in the second column So this means either the animal is yeah is near

the corners or near the walls but not the corners or near the edges of the of the box So we did the first two lines and we said yeah so when the animal is near the corners it usually it

usually the speed of the animal is usually less is usually lower right so this is actually this can correlate with animal speed and maybe we're just keeping like taking the animal speed information to the call this and it's actually this is actually true for the first two lines

where if we see on the far right where we see the results for all animals we see that if we use the speed as the decoder So the blue, the blue bar And the classifier So the classifiers I explain in the in the previous slide The frame or the coding framework It's not very different Right

Maybe the classifiers are a little bit better But it's it doesn't it's not really significant But if we take both the edge The edge of the box So combine the walls and the corners There actually the speed cannot predict this anymore The speed only

And we actually need more information than that when and the functional ultrasound

data gives us this information.So we could also train a decoder for animal speed in a

set of animals and test it on a different animal So all that I showed on the on the previous slide is trained on a session or a set of sessions for a given animal and then test it on a set of animals and then using another animal train and test it in another animal

So we did this by using a voxel signals but averaged air rise signals since we did we did not have exactly the same plane the same imaging plane for every animal but we had planes that captured hipocampus and tetraglambas and regions that are important for the decoding part

So we said okay so let's for the inter animal decoding let's try taking average rois signals training them training the decoder on them and then trying to test the decoder on the

same regions but in a different animal and so here on the on the top right the train where

we show here a training animal and a validation animal we see that they are different but they are not the same imaging plane but they present most of the same regions in this way in the in the popocamp in the doppler frame here on the on the on the top right the train where we

want to translate the weights from the same region across regions in this patient's aggression system So on the bottom left Here we see So yeah, on the bottom left We have the prediction versus actual speed And we see that So the The The The Demi The Demi

Decoder is in green And We see that it cannot really decode Decode anything Regardless of the actual Of the actual speed Then in orange we have the intranymal decoder that it's a little bit better And we also compare it to the intranymal decoder So trained on the same animal and tested on the same animal So obviously The The The intranymal

decoder is better But We show that the regressor in the The coder in the Interanymal Decoding is still Above chance when compared to To a dummy model

Okay, so we're moving to a different section here in collaboration with Sir Sebastian Castedo Simona Coco and Revi Monason from the Cubio Lab in ANS We tested a computational model that they were already working on against our data set

So the goal here was to model energy consumption of the spatially tuned neurons during spatial navigation So the assumption we make here is that when the animal moves There is a bulk There's a bulk of activity coming from the firing rate of the activity of the active neurons

And if the animal moves faster This bulk of activity will increase So the energy consumed by the energy consumed by the system will be the integral or And will increase quadratically with animal speed Because if the firing rate is linear to speed

The integral of it will be quadrat quadratic to animal speed So the model that we are looking at links energy and speed And we showed in our results that we have a strongly link between cerebral blood volume and speed So the assumption we make

is that energy the energy we are looking for here can be seen as a proxy as a proxy of the rate of oxygen consumed by the brain CMRO two, and from literature we have this link between CMRO two and CVV that is CVV The link is cbv squared actually

Same row two links to cbv squared And so the model in the end we are looking at is cbv squared equals a plus b times the speed of the animal power to the alpha

And we are proposing that alpha here is two for the link between speed and energy quadratic So alpha here is two And b obviously is non-zero

otherwise it wouldn't matter right So b is non-zero and alpha is near two So we can actually find alpha around two for the animals that have a very good decoding analysis

on the on the decoding part that I showed before And here we see the actual data We took the cbv data from the hippocampus So that's what we see in orange So this is the cbv squared already And we fit this to the model

And then what we get is the green is the green data So this is the fitted model And using only the speed signal Right So only this with only the speed signal We fit the model and we get the green curve The green line So in this plot It's a little bit more convoluted

But what we have is on the big plot we have d versus alpha For different animals and in the inset on the top right We have we actually have the error versus alpha plot And both need to be analyzed together because as I said before We want to so we want to minimize the error But as I said before

The B should not be 0 Right B should not be negligible Otherwise alpha can be anything So we can see, we can see the minimum being reached for the alphas around 2

For the Q versus alpha pot And for this, so yeah And for these animals We see that B is 0 So we see the alpha near 2 Which are the dips that we see on the Q versus alpha pot And on the B versus alpha pot We have this vertical lines that show where the dip is

And we see that the dip here corresponds to B to B being 0 or non-negligible So the last thing we have in the paper And I want to show here Is that we used our data to look at the results of a recent study that today is now

is not that recent anymore But it came out in 2023 That presented minute scale frequencies in the medial interrural cortex As it could be some sort of matronome for this patient aggregation in neurons in the medium term of cortex So they first looked at the autocorrelation

of the firing rate of these different neurons And that's what we see on this image The raster plot here And they find this oscillation pattern that are at minute scale frequency And if they look

And then they look at the frequency content of this or the frequency content looks like this And we have this minute scale oscillations from around 0.05hertz to 0.02hertz That goes from 60 seconds to 200,300 seconds Of period

So we wanted to look in our data and see if we could see something similar in the in the in the CBV data Here we see the autocorrelation similar

to what they did with neural activity But here we do with CBV with the CBV data The autocorrelation in different in different regions So paricopampus in red Then hippocampus in purple And somatosensory cortex And cadet putamen

And we see that these autocorrelations are yeah So these are entire regions Right So these are the average the average autocorrelation of these regions Sorry, so we see So this is on the first line We have the voxel the voxel versus the autocorrelation

So each line is a is a different voxel on that region And what we see on the bottom is the is the power spectrum density of these autocorrelation And we see that the average of it on color like the color of line

And each line is a is a different voxel on that region So we see that so here it's on the pocampo the pocampo region We see this oscillation pattern quite vividly And we see the peak

Also in the parse back and density around zero point zero five hertz A little bit on the para On the para hippocampus And faintly A little bit less Some voxels do present them But on average We have the other regions that do not present This minute scale peak

So this is for Yeah So we also present As we have this information for each voxel in the in the in our imaging plan We can see that We can see that in maps like this one So we have instead of looking at the parse back and density for a given voxel We can construct maps for

each being each being one We can see on zero point zero five hertz We see that in the hippocampi region We have this peak And that's the peak that corresponds to the purple The purple peak that we see on above

So for yeah and this is for a given animal for an example And if we do this so we search across different sessions Different imaging sessions And we what we wanted to understand is that

if this minute scale fluctuations are intrinsic or it like does it have anything to do with speed Right We saw the correlations with speed that were quite strong So we wanted to see if this if this peaks in the power spectrum density would have something to do with the animal speed

So what we did is we applied the same the same analysis for the voxel time series The voxel the cvv time series to the animal speed So we have the auto correlation of the animal speed And then this what we what

we see here in black is the parse spectrum density for the correlation of animal speed And we see that they align quite well the peaks align quite well with the with the peaks that we find on the parse spectrum density of the autocorrelation of the cbv data on the hippocampal

formation and the parahippocampal region So our understanding here is that the cbv the minuscule oscillation that we see in the cbv aligns with actually the animal speed fluctuation So basically when the animal starts to move or not in a very low frequency rate

If you have like a low pass filter on the animal speed doing the experiment So in summary we have a unidirectional information flow of vascular activity during the spatial navigation within regions that are very important for the spatial navigation system going from

dorsal thalamus to the parahippocampal region to finally the coding but a little bit weaker but we can also do inter animal decoding and we also decoded animal speed via interpolation and I'm

sure this can be further improved we showed a computational model for energy consumption during spatial navigation that is consistent with the CBV in the hippocampus for individual animals.

For demo purposes that we could get very good correlation with animal speed and we also show that there's low we have low CBV oscillations in the hippocampus

And these are aligned with the minuscule speed fluctuations that we find in our data As well So, thank My co-authors, Everyone that made this paper possible This project

And I thank you for listening And happy to answer any questions Thank you, Felipe Thank you for this great presentation It's very interesting So, to anyone attending Feel free to type your questions in the chat I will read them to Felipe And while you start typing

We can actually start with some questions that we received through the registration form So, In particular We were asked a bit more information about your animal handling protocol and how much time you take from let's say the surgery

through training up to the experiment Yeah, so this definitely these

are not like fast experiments So I would say from like from when the animal

arrives to the to the lab to the first experiment It's at least at least five weeks right because then because we would have like one week of abituation to the environment Then craniotomy and post a better aperitive care that would last for a week then abituation that

would also last for a week so yeah so at least at least a month for the experiments themselves And for the experiments We would have it would vary a little bit on the animal It's not so the craniotomy is quite a big surgery

So it depends also on how we can so the we're quite good now But it's a hard surgery And we would have from yeah

We could get animals up to two months like having a reliable signal on the brain And yeah, up to two months I would say so in total Yeah, maybe you can have like up to three or four months with the same animal from the beginning

From like when it arrives to lab to the last to last experiments Thank you So we have one question in the chat saying hey Felipe Thank you very much for the nice presentation How do you do with motion artifacts if any in functional ultrasound

Do you use some corrections for the aberrations induced by the skull Well, the skull is removed So I guess skull aberration are not But yeah, What about the exact effects Yeah, so So this is something that

So I do not have data to confirm this But from my experiment looking at had fixed experiments on awake animals I do think that freely moving experiments

on rats at least I'm not sure if on ma on mice would be the same But on rats we have less motion artifacts due to the animal movement Because what we do so we have a mohah helmet that is attached to the animal to the animal's head

And then this helmet is where the probe is fixed right So this was also part of my PhD like handling developing this helmet and improving the improving on this helmet And if you have something

that where the probe is quite secure And you are attaching this to the animal's head I think it kind of follows the animal on the movement It does, and it's not so this where I compare to the head fix setup where the

animal can do like the forces it do It's directly it's directly touching the probe right Even though of

course it's like a solid a solid setup But the forces are being are opposite So it can move But it's making some like when it tries to move It's making some force that is applied to the probe right And I think that on freely moving experiments

This is less the case because we did everything possible so that the probe and the probe holder follows the movement of the animal but so this said This means that I think I have less motion artifacts than had fixed

But we still have we still have some movement that simply doing like frame wise correction is enough for me Like it's I have a very good video after that and I still have then some frames that will have some motion artifact

And then I will I'll just take them out But this is always less than four or five percent so my threshold is removing five percent of the frames if they are artifacted

Perfect, thank you very much for this answer I hope ah have yeah we answered your question We are also asked if you ever looked at functional ultrasound data associated with discrete behavior So not like speed and navigation But more like freezing

Grooming, or any discrete stimulus like receiving drugs or stimulations Yes, so I am assuming you're not talking about like the usual event based simulation Right That you

do like simulation during a period of time And so you're talking really about like one stimulus and then seeing what happens So yes, it's the tricky part So I think there's So there are there are papers on freezing

here from the lab that talk about this But I do not take part on those But we did some tests with like injection in injecting molecules or things like that It's not that different It might be a little bit

trickier when you have just one stimulus that you are really like touching the animal Or like injecting something or like really changing something Cause we will definitely have some abrupt movement of the animal And this is a bit tricky But usually this is

something that happens quite quickly And then with the nervous recruitment You would be able to like But I don't think the same the same analysis apply here Because the correlation this correlation something that is really done like it is very good when there

is a whole so your whole experiment matters right The from start to finish here you have of course your the whole experiment matters But you have the baseline And then after that you have some You don't know how long it will take for the stimulus to take place So I think for this

Like a better analysis would be looking at what really what happens to the voxels when you do your discrete stimulation And trying to go for it from there I don't think the correlation the correlation analysis will be the best I think it will give you something But I don't think it would be

the best thing to do in this case I think we have time for one last question So I'm picking one that I found interesting from the ones that we received during the registration part That if you can comment briefly on the pros

and cons of functional ultrasound in comparison to fluoration based in vivo imaging like miniscopes or other optical based methods Yeah, I think I think they are different right That there, are complementary in the sense that

we have a much higher like a much bigger field of view And also a less good spatial resolution compared to this techniques And temporal resolution So in terms of like deep regions in the brain

With fluorescent based imaging would be very hard Normally when people want to image that either

they should like either they take part of the of the cortex out right or they or they would like maybe like if you want to image like the part of a couple part is a little bit less invasive because then you can kind of get an angle there like to put a mirror and image the part of a

couple region but yeah if you want anything like a little bit deeper you would you would have to take part of the brain out and this is hard but yeah I don't I don't I think I don't see I think they are complimentary like they are they are not for the same the same questions I agree and we can

conclude with a last one from the chat canal say hi thank you for your presentation do you have any issues due to the cable being broken or starty and limiting the animal movements yeah so you

said mice but yeah so my experiments were in rats right and a little bit because of this as well Because so we We actually don't have the We don't usually do freely moving on mice I know like it's starting to be done

And people are like: 'Really taking care of this How will you handle the probe The cable and all that.' For rats is Is still not perfect movement Right It's still not a completely natural movement But it handles much Much better than mice from

Where I see Like we wanted to do expansion of agents We wanted to take like the movement of the animal to the extreme Right We wanted it to explore We wanted it to move And And And this I think is like kind of the extreme The extreme experiments to do with function of a sound with the cable and all that But there's much like much more like other

experiments that I think where the animal doesn't need to move that much Maybe while you're doing like learning to choose a button to press or things like this Where the animal is in maybe like trying with rats and seeing if you if you have super visible this

is super visible and I think even like even if we have mice but if you want really the animal to

move and do like yes I confirmed that I've seen both and definitely we can handle the cable and

they weigh much better than mice but still for mice I agree with felipe that maybe other kind of tasks on tea or wine mazes or linear corridors can be more feasible compared to like a special

exploration in to the arena where they can move everywhere and we can always discuss this further via email and same goes for all the questions that we couldn't answer from the registration form We will get back to you via email in the coming days and with this I wanna thank again felipe

very much for your presentation thank you to all of you for attending and participating