Cole Korponay “Pushing the boundaries of fMRI-driven brain feature discovery by leveraging neuroanatomy- and electrophysiology-informed computational approaches”

March 09, 2023

Information

ID: 9624
To Cite: DCA Citation Guide

Download Transcript

00:06Hi. For those of you who don't know me,
00:09my name is Cole Courtney.
00:10This is my first time that was there.
00:13Just want to thank Todd and the
00:15organizing committee for giving me the
00:17opportunity to share my work here.
00:19So let's dive in.
00:23I'd like to harken back to some of what we
00:25talked about during yesterday's session,
00:27namely how we can get past the kind
00:30of .28 correlations between brain
00:32and behavior that sage kind of
00:35mocked us thought in her bingo card.
00:38So we've talked a lot about different
00:40strategies for doing such a thing right here.
00:43Simon talked to us about more intelligent
00:46use of incorporating confounds into our
00:49analysis pot and genuine talk to us about
00:52different algorithmic choices we can make.
00:54Ruby talked to us about different
00:57parcellation strategies.
00:58We've touched on dense sampling
01:01tech advances that give us better
01:03spatial and temporal resolution,
01:05sample sizes, etcetera.
01:08So I think these all kind of more
01:10or less fall under the category
01:12of ways we can try to extract more
01:14signal from our resting state data
01:16or our functional connectivity data.
01:20And I'd like to approach this
01:21from someone of a different angle
01:22to talk a little bit more about
01:24the nature of the signal itself,
01:26and in particular the way we
01:28dissect and package this functional
01:30connectivity data to put it into
01:33our analysis as an input.
01:38So to be slightly hyperbolic and pejorative,
01:41I think a lot of our neuroimaging
01:44metrics and analysis come
01:46from this sort of top down.
01:48Here's a cool math thing we can do.
01:50Let's throw it out our data
01:53and see what happens. And.
01:55The idea or the notion of whether
01:57these methods or metrics actually
01:59model the underlying neural
02:01biology is almost an afterthought.
02:06And so I'd like to advocate for a more
02:08bottom up approach where we start with
02:10some brown truth features of circuits
02:12that we know about from electrophiles
02:14or no anatomy and design our metrics
02:17and analysis around these ground
02:19truth metrics to try to pick up their
02:21signatures in our neuroimaging data.
02:27And so I call these neuro anatomically
02:30and electrophysiologically
02:30informed neural imaging metrics.
02:35So let's walk through what
02:36that might look like,
02:37and I'll focus on corticostriatal circuitry
02:39because that is my area of expertise.
02:42So what are three things we know for
02:45sure about how corticostriatal circuits work?
02:47So one is that individual striatal
02:50nodes receive convergent input from
02:52lots of different areas of cortex.
02:54OK, second, we know that in order
02:57for striatal projection neurons fire,
02:59they need to receive temporally
03:01synchronous input from these
03:03different conversion inputs.
03:043rd we know that striatal projection
03:06neurons fluctuate between these
03:08three electrophysiological states,
03:10these two resting states,
03:11and then a burst firing state
03:13that happens rarely.
03:17So let's start with this first one that
03:20straddle neurons receive convergent
03:22input from different areas of cortex.
03:25So the main kind of take away
03:27from this insight is that.
03:29What really matters for shaping
03:31how astrild neuron is active
03:32and what its function is,
03:34is not the strength of any
03:36one input in particular,
03:37but more about the relative composition
03:40of these inputs and how they're balanced.
03:44Kind of shapes the input.
03:47And this is really at odds with
03:49how we typically do neuroimaging
03:50analysis on corticostriatal circuits.
03:52We look at these things A to B connections,
03:55the correlation between the time series
03:57of 1 cortical node and one straight node.
04:02Or, you know, a collection of these things,
04:04but which are treated independently
04:06in our analysis frameworks.
04:09And it kind of reminds me of like,
04:11OK, if we wanted to.
04:13Assess how an orchestra is doing.
04:17What information would we get
04:19out of just asking, OK, what?
04:20What's the average volume of the
04:23harpist through the song or whatever?
04:25That doesn't really tell us much
04:26without the context of how all the
04:28other instruments are playing, right?
04:30The whole point of an orchestra is to
04:32achieve this sort of synchrony of this
04:34instruments playing at this volume.
04:35This instrument is playing at this
04:37volume and we can only really
04:39assess the full ensemble by.
04:40Considering everything together.
04:42So here, OK, the harp is playing less
04:44loudly than the bass and the piano,
04:46but more loudly than the violin.
04:47OK, cool. It's doing something good.
04:49That's not information we could have
04:51ascertained just by looking at what
04:54the volume of the heart player was.
04:56And so kind of bringing this back
04:57to the cortical stradal circuits,
04:58I'd like to advocate for the use
05:01of the connectivity profile as
05:03kind of a fundamental metric for
05:05our analysis inputs over and above
05:08the the A to B connection.
05:10And so we can ask like different properties
05:13of this connectivity profile we can ask.
05:15What's the rank order arrangement
05:17of which connections are coming in
05:19most strongly versus less strongly?
05:21What's the aggregate strength
05:23of these things?
05:24To what extent are these communications
05:27delivering inputs in temporal synchrony?
05:29Are there different configuration
05:30states as we go across time and a scam,
05:33so different ways we can metricized this?
05:37And just a brief example of the
05:39utility of using this kind of
05:41connectivity profile approach.
05:43So this is from a recent paper
05:46in Neuropsychopharmacology.
05:47We're looking at smokers versus
05:49non-smokers and looking at differences
05:51in cortical striatal circuits
05:53between these groups.
05:55And so we found this interesting
05:57dissociation whereby in the dorsolateral
05:59striatum we find evidence of rank what
06:02we call rank order disarrangement,
06:05so compared to.
06:06Here the order of connection
06:09strengths that are non-smoker.
06:11These are all cortical inputs to one example,
06:14strike of voxel.
06:14This is the strongest input down
06:16to the least strongest output.
06:18If we compare that to smokers,
06:20they're totally swamped around.
06:21And what's interesting is that
06:23the total strength,
06:24if you add up all of these and
06:25add up all of these,
06:26those aren't significantly different.
06:28Just the order of them totally swapped
06:31around and we only see that phenomenon
06:33in the dorsal lateral trade up.
06:35Conversely,
06:35we see kind of the opposite and
06:38the ventral and medial striatum,
06:40we don't see any of this arrangement of
06:41the rank orders of connectivity profiles.
06:43What we instead see is preserved rank orders,
06:46but what we call aggregate divergence
06:49of the connectivity profile where
06:51everything is just proportionately stronger.
06:54And so this is just to say that
06:56these are insights about circuit
06:57pathology that are more nuanced
06:59than we could ever have gotten with
07:01just A to B connection framework.
07:03We get to learn more about how the
07:06circuits are disrupted and how those
07:09differ in different spatial areas.
07:14As one more example,
07:15so I mentioned that striatal nodes
07:18fluctuate between these three
07:20electrophysiological states.
07:21They spend most of their time toggling
07:24between a down hyperpolarized state and
07:27up slightly more depolarized state,
07:29and then occasionally these
07:32spontaneous versus viruses.
07:36And so if you wanted to try to
07:38recover a signature of that
07:40phenomenon in resting state data,
07:42I think the most obvious thing
07:43you could try it first is just OK.
07:45Let's K means, with K = 3 the striatal
07:47bowl signal of individual striatal boxes.
07:50What do we see a pattern like that
07:53that reproduces those properties?
07:55Do we see something like that?
07:57The answer is no.
07:59You just see kind of the reconfiguration
08:02states that are occupied about
08:04equal amounts of time and it's not
08:08recapitulating what we see at the
08:11underlying electrophysiological level.
08:12OK, so how else might we search
08:14for these states?
08:15So again,
08:15let's use the kind of ground up approach.
08:17Let's consider what dictates these States and
08:21the answer is the stratums excitatory input.
08:23So.
08:24When striatal neurons are in
08:26that hyperpolarized state,
08:28it's because they're not receiving a
08:30ton of excitatory input from their,
08:32you know,
08:33cortical euthymic inputs when
08:35they reach that option.
08:36More depolarized state.
08:37OK, now there's some more synchronized
08:39input coming in from those excitatory areas,
08:42and then in those burst firing states,
08:43you're getting lots of convergent
08:46and temporary temporally synchronous
08:48activity from all of those places.
08:50OK,
08:50how can we use that knowledge
08:52to help model this stuff?
08:54Well,
08:55I think Rick Vessell yesterday was
08:57just talking to us about a really
08:59interesting way that we can actually
09:01measure when different edges are
09:04communicating in similar temporal patterns.
09:08So combining Rick's edge centric
09:12analysis method with the idea
09:14of connectivity profiles,
09:15we can do the following.
09:17So if we just take a stride of
09:19voxel here in the red and all of its
09:21cortical inputs and so on the left and a,
09:24those are just the the nodal time
09:26series of each of those ROI.
09:28And we use Rick's method to
09:30create these edge time series,
09:33these instantaneous coal fluctuation
09:35time series that's here and B.
09:38So we have eight of them, one for each edge.
09:43What we can do is then at each TR,
09:45and this is just a three example TR's,
09:47we can pull out a instantaneous
09:51cope fluctuation intensity profile
09:53which is just telling us what
09:56is the configuration of.
09:57Equal activation with portable
09:59inputs at any given moment in time.
10:02And if we cluster these,
10:03maybe we'll see something interesting.
10:08So again I use K means of three
10:10because that's what we know from
10:12the underlying neurobiology.
10:13OK, so we first see that,
10:15OK, we get three states.
10:18One in Blue is a state where striatal
10:20neurons are have negatively coordinated
10:22activity with their portable input.
10:25So basically the cortex is doing one thing,
10:27the striatum is doing another,
10:29another thing,
10:29it totally unrelated that kind of
10:31corresponds to that hyperpolarized see.
10:34Configuration 2IN green.
10:35OK, now there's some more coordinated
10:37activity with all of the cortical inputs
10:40that's reminiscent of that depolarized seat.
10:42And then in red,
10:43we're seeing really strong input
10:45from all of the cortical inputs.
10:47Now the question is,
10:49do these configurations display the
10:51temporal properties that are reminiscent
10:53of the underlying electrophysiology?
10:56And it seems that yes, maybe they do.
10:58So we're seeing that these states,
11:01this is an average across
11:04all striatal neurons,
11:05all serial boxes rather they're
11:06spending most of their time
11:08fluctuating between configuration,
11:10running configuration two and
11:11then every so often it's showing
11:13that first firing and we can
11:15see that quantified here.
11:17Spending much more time in configurations
11:19one and two than configuration period.
11:23So I hope I've convinced you
11:25of the following things so far,
11:26which is that these this bottom up neural
11:29anatomically and electrophysiologically
11:30informed or imaging metrics can one
11:33provide more nuanced insights about
11:35circuit pathology and disease States
11:37and potentially help us recover
11:40signatures of fundamental circuit
11:42properties in resting state F MRI data.
11:48OK, so what's next?
11:491st is we'd like to see if we replace
11:51those A to B connection inputs with
11:54connectivity profile property inputs and
11:57connectome based predictive modeling or
11:59other predictive modeling frameworks,
12:01do we get better predictive accuracy?
12:03That's One Direction.
12:06And the second is, you know,
12:08by finding signatures of these underlying
12:10circuit properties in neuroimaging data,
12:12we can do this thing which I call like
12:15scale up investigations of those properties.
12:17Usually when we investigate
12:19electrophiles or neuroanatomy,
12:21it's in a sort of spatially confined
12:23area and it's usually in an animal model.
12:26But if we can find signatures of that
12:28stuff in our human neuroimaging data,
12:29we can look at this in humans and
12:31at scale and at the systems level
12:33and the whole brain level.
12:38So thanks for listening in.
12:39I hope this was interesting.
12:42Big thanks to my mentors and
12:44collaborators as Dan Haber,
12:45Scott Lucas, Andy James,
12:47Elliot Stein, Tom Ross at Nida.
12:48If you want to follow updates on this
12:51work and get some bad MBA takes,
12:53you know where to find me.
12:54Thanks so much.
13:05Two questions.
13:08Station so.
13:11Matrix isn't every role like.
13:17It is but. So it it would actually be
13:20somewhat trivial to reconfigure that to
13:23to have it be the connectivity profile
13:26be the metric of analysis because.
13:29The metric right now are the cells in there,
13:31like each cell is an A to
13:33B connection metric.
13:34But we can just take the whole
13:36row instead and ask about the
13:37properties of the row.
13:38So it would be trivial to kind of.
13:40Attempt to do that,
13:41so it could just be interesting.
13:48But I guess my second question was more.
13:52I like this composition.
13:54Trying to build something more
13:57kind of warm when we're on live,
13:59more motivated and and that's
14:01really funny because at this
14:03level looking at this rate or.
14:08Build that out to a network.
14:13I just wonder if you have any.
14:15No, it's a great question.
14:16And like the properties that I
14:18described about corticostriatal
14:19circuits are going to be different
14:21in other circuits of the brain.
14:22And I think it's just going to
14:24require people who are experts in
14:26those various systems to build out,
14:28you know, that pipeline.
14:29But then of course the the problem
14:31you encounter is you don't want this
14:32patchwork thing and different systems.
14:34So yeah, that's just going to
14:35require some some bigger conceptual
14:36thoughts on how to get that done.
14:38So great question.
14:40Could you go back to the vessel?
14:46Photo of you.
14:51No one's going to use that little.
14:54That raised single speed. One question.
15:00Their message?
15:04Sorry smokers.
15:15Forward.
15:26We haven't looked at the the dynamic
15:28stuff when we smokers yet. Yeah.
15:30So that's we'll look at that next, yeah.
15:42But.
15:50Full scale with single neuron firing.
15:55Right. I mean. Though they're not happening
15:58at the same spatial scale, of course,
16:00although we are close to the temporal scale.
16:02This is HCP data at 7:20 millisecond TR's,
16:07just close to what's happening. But.
16:14Yeah. So the, you know, the next step
16:17here is to collaborate with people who
16:19do simultaneous F MRI and, you know,
16:21electrophysics recordings and animal models.
16:23And then we can really get a more
16:25robust sense of whether these are
16:27actually reflective of those properties
16:29because right now we just have,
16:30hey, these look like those things,
16:32but we need to go a little bit
16:34deeper to see if there's actually
16:36the causal relationship there. Yeah.