Rick Betzel “Edge-centric connectomics”

Transcript

• 00:07So really excited to be here today and
• 00:09talk about some work from from my lab.
• 00:11It's kind of unfold over the
• 00:13last two or three years.
• 00:15They try to tell you a a complicated
• 00:17story maybe not that complicated
• 00:19but one that's trying to address
• 00:21a very specific question,
• 00:23the very end I'll try to present
• 00:24a resolution to the question but
• 00:26I think the more interesting part
• 00:27is what happens in between like
• 00:29the journey to the resolution. So.
• 00:32So I'm going to do that right now,
• 00:34the talks inside the.
• 00:36Essentrics connectomics.
• 00:37So this is kind of a silly thing
• 00:39to say in this room,
• 00:40because everybody who knows this,
• 00:42but bringing networks really composed
• 00:43of two things, nodes and edges.
• 00:45Nodes of course represent populations
• 00:48of neurons, sensors, areas,
• 00:51and then the edges represent the
• 00:53functional anatomical connections
• 00:54between pairs of those those nodes.
• 00:56And I'm going to argue that
• 00:59network neuroscience,
• 01:00maybe neuroscience in general,
• 01:02has really been interested
• 01:04in properties of the notes,
• 01:06whether it's the number of connections
• 01:08they make.
• 01:08The number oh,
• 01:09sorry,
• 01:10the community to which a node is
• 01:12assigned centrality of a node with
• 01:14respect to a process most of our of
• 01:17our of our favorite measures are
• 01:18really related to to the nodes themselves.
• 01:21Those parallels has been going
• 01:23on in neuroscience for like the
• 01:24last century going back at least
• 01:26to broadman and have chopping up
• 01:27with the brain the basis inside
• 01:29work at the comic properties,
• 01:31but also including more more recent
• 01:33imaging based approaches for for
• 01:36for characterizing properties
• 01:37of areas and territories.
• 01:39I think that's great,
• 01:40but I think it means we're possibly
• 01:42leaving something on the table
• 01:44and we might benefit maybe from a
• 01:46shift in perspective, for instance,
• 01:48one that prioritizes other features
• 01:50of the network with nodes themselves,
• 01:53and there's precedent for how to do this.
• 01:56Maybe not so much in the neuroscience
• 01:58literature or human imaging literature,
• 02:00but if you go back to network science papers,
• 02:03this is probably the most famous one.
• 02:06But there are examples of how
• 02:07to shift that perspective to go
• 02:09from nodes to edges,
• 02:10and essentially what this
• 02:12paper shows is it presents a.
• 02:15Really,
• 02:15really simple strategy for taking
• 02:17our familiar node node networks for
• 02:20areas connected by structural or
• 02:22functional connections and flipping
• 02:24them so that the new nodes or the
• 02:26edges in the original network,
• 02:27it's a, it's a, it's a.
• 02:30I just don't know. I don't think.
• 02:33I don't think I touch anything.
• 02:35It might be a battery.
• 02:41Oh, we have batteries. That's great.
• 02:44Sorry, everybody. How they're
• 02:45going to go skiing at 11:30.
• 02:47Yeah.
• 03:00Way to go, Rick. Yeah. Back.
• 03:09We'll see.
• 03:12Hey, I hear you again.
• 03:16Well, that's working excellent.
• 03:19I got a bunch of devices like this.
• 03:26Enjoy it here.
• 03:27So you're saying this is an approach
• 03:30essentially transforms node node networks,
• 03:33ones we're used to dealing with,
• 03:34into a kind of higher order network
• 03:36with the new nodes or in fact the
• 03:38edges in the original network.
• 03:40I won't go into details of how to do
• 03:42this spell present just to kind of a
• 03:44high level schematic about my work.
• 03:46So in this particular approach the
• 03:49strategy is let's grab 2 edges E1E2.
• 03:51They must have a shared stub,
• 03:53so one of the endpoints must
• 03:55be coming to both.
• 03:56This leaves 2 unpaired or unmatched.
• 04:00Angulate the overlap in the
• 04:01in their connectivity profiles
• 04:02and that gives you one number.
• 04:04In this case there's three edges
• 04:06and overlapping by back by 1:00,
• 04:08so you got an overlap measure of 1 / 3.
• 04:13Do this for all edges and now you
• 04:14get a new connectivity matrix.
• 04:16The rows and columns correspond
• 04:18to edges in the original network.
• 04:20The entries correspond to this weighted
• 04:22overlap with connectivity profiles.
• 04:25So very simple.
• 04:26Some advantages this approach.
• 04:28So for instance,
• 04:30if we're closing this network,
• 04:31you're clustering edges,
• 04:32you're not clustering areas or parcels,
• 04:34and that means the edges inherit community
• 04:37assignments or module assignments,
• 04:38and the perspective of any node
• 04:40will it inherits the communities
• 04:42of its edge distance.
• 04:43They overlapping, which is kind of cool.
• 04:47There's at least one other
• 04:48approach to doing this,
• 04:49but in the end it generates
• 04:51something you know awfully similar.
• 04:54So why don't we just take these off the
• 04:56shelf and start applying them to brain data?
• 04:59There's some challenges.
• 05:00First, these two approaches really
• 05:03only work for sparse networks,
• 05:06meaning most of the connections
• 05:07are absent or not present.
• 05:09Doesn't do a good job of dealing
• 05:12with signed connections.
• 05:13And hey, we we we like correlation
• 05:15matrices and there's negative correlations.
• 05:18And that's where everything is and kind of
• 05:21kind of like and tested territory maybe,
• 05:23but it deals with grass and self,
• 05:26transforms the networks,
• 05:27the static networks into edge edge networks
• 05:30and so you lose all temporal information.
• 05:33So maybe that's not ideal.
• 05:36So back in the before times like 2019 or so,
• 05:39we sat down and started thinking about
• 05:41what an edge centric approach for your
• 05:43imaging and brain data might look like.
• 05:45There's a bunch of people involved
• 05:47in this project and we decided to
• 05:49start with something that we were
• 05:51already eminently familiar with and
• 05:53that is functional connectivity
• 05:55has defined as a correlation.
• 05:57Again, this is really probably
• 05:59don't need to say this.
• 06:00I don't know.
• 06:01Spell it out again.
• 06:02The idea is records from two parts of the
• 06:04brain and calculate the shared variance.
• 06:07Correlation coefficient becomes
• 06:08a weight in the matrix.
• 06:10There's our functional connectivity.
• 06:24So I actually want to try to
• 06:26unpack that stuff a little bit.
• 06:28This is really simple stats unpacking,
• 06:30but I'm going to do it anyway.
• 06:32So what do we actually do?
• 06:33We calculate a correlation.
• 06:35What? We take those time courses
• 06:37from two parts of the brain.
• 06:38We start by Z scoring 0 mean,
• 06:41unit variance, and we start
• 06:42calculating a bunch of products and
• 06:44each instance in time we calculate
• 06:46the product of these two time series.
• 06:48This gives US1 number a an
• 06:52instantaneous Co fluctuation.
• 06:54It's signed.
• 06:54It's amplitude tells us both of those two,
• 06:57the blue and the Red Time series are
• 06:59moving in the same direction with
• 07:00respect to their mean and by how much.
• 07:02We repeat this for all frames
• 07:04and it gives us.
• 07:06Well, gives us a Co fluctuation
• 07:08value at every instant in time.
• 07:10The average of that is correlation.
• 07:13That's our our functional connectivity.
• 07:16What if we did something a little devious?
• 07:18What if we like still admitted that
• 07:20last step, the averaging step?
• 07:22Well, goodbye correlation.
• 07:24Goodbye functional connectivity
• 07:25by arguably preserve something
• 07:27that's quite useful.
• 07:29And it's this cooperation time series, right,
• 07:32we're actually preserve that temporal.
• 07:35And this cool fluctuation time series
• 07:37has an interesting properties.
• 07:39It tells us about the amplitudes,
• 07:40so how big the come fluctuations are.
• 07:42It tells us about the balance.
• 07:44Like those two time series are up
• 07:46and down together and telling us
• 07:48exactly when in time those cook
• 07:51fluctuations are happening.
• 07:52So, for instance, here's a.
• 07:55There's a big conflagration here.
• 07:58Why is it big? Why is it positive?
• 08:01The red and the Blue Time series both
• 08:04have big Z score values at that time.
• 08:07Here's a negative Co fluctuation
• 08:08wise is negative ones going up,
• 08:10the other one is going down.
• 08:13You get one of these for every pair of
• 08:15brain regions and every time you calculate
• 08:17functional connectivity is a correlation,
• 08:19be a full or lag or partial.
• 08:21You're implicitly doing what
• 08:23doing this step exactly.
• 08:24You might be averaging at the end,
• 08:26but you are calculating the
• 08:28Co fluctuation time series.
• 08:30And my clean I'm gonna pack over
• 08:32the next few slides is that I think.
• 08:36These go fluctuation time series
• 08:38have some interesting and potentially
• 08:40useful properties.
• 08:42So what do we get when we calculate this?
• 08:44Well, rather than doing it for a single node,
• 08:46or rather for a single pair of nodes,
• 08:48most of which were all pairs
• 08:50of nodes for all edges.
• 08:51So that time series I showed you in
• 08:53the previous slide, it's like a slice.
• 08:55This matrix rows here are edges
• 08:58and columns are time.
• 08:59But if we slice this vertically,
• 09:02what do we get?
• 09:03Well, now we're aggregating
• 09:04a collections for all edges,
• 09:06all pairs of brain regions,
• 09:08and we're getting time resolved
• 09:10conflations and resolve networks.
• 09:12There's no sliding window,
• 09:14there's no kernel, no convolution.
• 09:16We just get instantaneous conflation that
• 09:20go fluctuations. That works for free.
• 09:23There's also some interesting properties.
• 09:25Remember code fluctuation.
• 09:26Their average is a correlation
• 09:28coefficient collapses across time.
• 09:30I get other factor of correlation
• 09:32coefficients. Rearrange it.
• 09:34Upper triangle of a square matrix.
• 09:37There's that. C again.
• 09:38So this is.
• 09:39This is an exact decomposition
• 09:41of functional connectivity into
• 09:43time resolved to fluctuation.
• 09:45We're getting networks for free.
• 09:47And it's giving us these frame wise
• 09:49estimates of the numbers, yeah.
• 09:52So we started looking at these
• 09:54cool fluctuation time series and
• 09:55we noticed them what we think are
• 09:57kind of interesting properties.
• 09:59They all seem to have these periods of
• 10:02quietude punctuated by these big bursts.
• 10:05You can see some here.
• 10:07Now remember the fluctuation time series,
• 10:10the products of activity,
• 10:12the average of them is functional
• 10:15connectivity.
• 10:16That means those high amplitude
• 10:17frames that kind of tipped the scale,
• 10:19they took that average a little bit.
• 10:21They contribute more to that.
• 10:22Average pattern and so we asked
• 10:25do those bumps,
• 10:27those high amplitude equal fluctuations
• 10:29they tend to occur synchronously
• 10:31across edges and think that maybe
• 10:34brain wide events or do they occur
• 10:37asynchronously just kind of uncorrelated way?
• 10:40Here's that same.
• 10:43Whole grain matrix.
• 10:44If you're looking at it,
• 10:46maybe you already know the answer to this.
• 10:48This is kind of like vertical band
• 10:51the striations.
• 10:52Those are frames and time instance
• 10:54and time when lots of edges
• 10:56simultaneously have big Co fluctuations,
• 10:58positive or negative.
• 10:59You can see this when you look at the
• 11:03global amplitude and highlighting a couple.
• 11:05If possible,
• 11:06the pure putative events,
• 11:08right?
• 11:09The distribution for those
• 11:11amplitudes is heavy tailed.
• 11:13And you just didn't really
• 11:15drive this point home.
• 11:16Here is the network,
• 11:18here is the network structure during
• 11:20some of those flammable 2 points versus
• 11:23the low amplitude strikingly different.
• 11:25And this to us suggested
• 11:27that the network dynamics,
• 11:28the way the network evolves over time,
• 11:30it's kind of bursty.
• 11:31It goes between these periods of high
• 11:34amplitude and relative low amplitude.
• 11:36I want to be clear, if we're not the
• 11:38first people to think about this,
• 11:39other people have written a lot about this,
• 11:42and that's kind of a.
• 11:44Again, my *** handed to me on Twitter
• 11:46because this was what happens.
• 11:48But it's. But it's true.
• 11:50But I'm saying here that there's
• 11:52some advantages to our approach.
• 11:53There's an exact composition, parameter free.
• 11:56That's the whole brain.
• 11:57There's some reason why we might like it.
• 12:00So one of the implications
• 12:02then of this like you know,
• 12:04potentially burst the Co fluctuations.
• 12:08The implication that becomes
• 12:09clearer when we look at the the.
• 12:12The tales of the amplitude distribution.
• 12:15So essentially what we're doing is
• 12:16we're taking each frame for filter,
• 12:18we're filtering them.
• 12:19We're retaining just the top
• 12:21amplitude and the lowest amplitude.
• 12:23The left is the top and the
• 12:26bottom is the right.
• 12:27If you just look at this.
• 12:30I'm asking you to look at it.
• 12:32Left hand side looks like
• 12:33functional connectivity.
• 12:34Right hand side is kind of paler.
• 12:36The Co fluctuations are very weak.
• 12:39And in fact,
• 12:40if you calculate the correlation of
• 12:42each side of this matrix with I'm
• 12:44average static functional connectivity,
• 12:46the high amplitude friends are much more
• 12:49strongly related versus the bottom.
• 12:50They also have higher modularity,
• 12:52so there's a stronger modular system level,
• 12:56organization, organization,
• 12:57and I amplitude frames than in the lower.
• 13:01This adjustment by percent of the
• 13:02frames you can do with, you know,
• 13:04really any percentile you like.
• 13:06And it suggests to us that functional
• 13:08connectivity can bring systems.
• 13:10We can reasonably explain them on
• 13:12the basis of these rare network
• 13:14wide dude go fluctuations.
• 13:16So we started asking,
• 13:17well,
• 13:18what other properties these these
• 13:21events have?
• 13:23And this is kind of a hodgepodge of results.
• 13:25You got a Mile High view of it.
• 13:28So we looked at what's happening in
• 13:30terms of brain activity during the events.
• 13:32That's really dominated by
• 13:33a single mode of activity.
• 13:35Depending on where you're coming from,
• 13:37this might be extrinsic versus
• 13:39intrinsic division might be
• 13:40the first principle gradient,
• 13:42it might be the default mode,
• 13:44but it's a very recognizable
• 13:46pattern of activity.
• 13:47The events themselves have
• 13:50synchronized during movie watching.
• 13:52They're not entirely intrinsically driven.
• 13:54They are partially dependent
• 13:57upon the stimuli.
• 13:59They actually used to be
• 14:00reconstructing networks for the
• 14:02events versus the low amplitude.
• 14:03They lead to stronger brain behavior,
• 14:05correlations and enhance brain fingerprints.
• 14:08This is with scan club data.
• 14:11We ask can we identify
• 14:13individuals across scans?
• 14:14Expectation is that within individuals
• 14:16we'd see a stronger correspondence
• 14:18and. Between the sea attenuated
• 14:20similarity and we see this.
• 14:23So the high amplitude this is the bottom.
• 14:25But this is subject by
• 14:28subject similarity matrices.
• 14:30And again, as I said,
• 14:32this enhances brain behavior correlations.
• 14:35This is CP data.
• 14:3810 behavioral factors and that colored
• 14:42points here represent the excuse me.
• 14:52Sometimes it could take a breath.
• 14:57Yeah, the colored points represent the
• 15:00correlations obtained from the high
• 15:02amplitude and the grave and below amplitude.
• 15:04We follow this up with a few other studies.
• 15:07Events themselves are not monolithic
• 15:09and of correspondence distinct states.
• 15:11The propensity for event to occur within a
• 15:14within a scan session is is that correlated
• 15:18with endogenous fluctuations and hormones.
• 15:20We've been a biophysical models
• 15:22that partially explain where the
• 15:24events are coming from.
• 15:26This model take structural connectivity
• 15:29couples brain regions together the treat each
• 15:32as oscillator stimulates the new data and we.
• 15:35Events in these data,
• 15:36when we destroy the structures and
• 15:39specifically the modules events go away.
• 15:41It was telling us that the underlying
• 15:44modular structure of SC might play a
• 15:46role in the emergence of the events.
• 15:49And something we're still kind of puzzled by.
• 15:52Going back to some of the
• 15:53movie watching data,
• 15:54we found that the timing of events
• 15:56is also very stereotypical.
• 15:58And these red lines or spond to
• 16:00the ending of movie clips and
• 16:02that's when everybody has an event.
• 16:04But here's a big burst at the end of movies.
• 16:07We're trying to kind of figure
• 16:09out what's going on there.
• 16:12So I'm going to end with the
• 16:15other couple slides, but like,
• 16:16I want to make a couple of points
• 16:18and have some caveats.
• 16:19I've really focused on the high
• 16:21amplitude frames, these big events,
• 16:22they're really, really part of the story.
• 16:26For instance, uh,
• 16:27the identifiability or that fingerprints.
• 16:30They actually don't tend to people the
• 16:32highest amplitude with the next bin down.
• 16:33Other people have shown this as well.
• 16:36I think this has a lot to do with how
• 16:39events themselves are related to one another.
• 16:42I have two friends are
• 16:44very stereotypical events.
• 16:45Go fluctuation patterns are
• 16:46very similar to one another.
• 16:48So the.
• 16:52So when you look at that highest bin,
• 16:53you're essentially getting lots
• 16:55of the same pattern.
• 16:56You go down, you get a mixture.
• 16:58Something looks a little bit more
• 17:00like functional connectivity.
• 17:01And I don't think that global
• 17:03events are the full story.
• 17:04So far we've been looking at the
• 17:06whole grain go fluctuation signal.
• 17:08You can calculate Co fluctuation
• 17:10time series for each system and
• 17:13they are differentially coupled
• 17:14to that whole grain signal.
• 17:17Those system level Co fluctuations
• 17:19also have their own.
• 17:20Your potentially interesting
• 17:22correlations director, too.
• 17:24And then there's just kind of a.
• 17:27Maybe elephant in the room in some ways.
• 17:29Since we published our first paper,
• 17:30there's been at least a handful that
• 17:33have claimed that some of the apparent
• 17:35temporal structure we're seeing,
• 17:37you know,
• 17:37maybe it's not meaningful or maybe
• 17:40my meaningful is that if you were
• 17:42to take a static correlation matrix,
• 17:44there's no temporal structure,
• 17:46use it to generate simulated data.
• 17:48The simulated data inherits some of the
• 17:51properties that we see in the real data.
• 17:53I think these are really good studies.
• 17:55So they present a possible.
• 17:57Mechanism for when these events come from.
• 18:01It's to me it's not particularly
• 18:04satisfying explanation.
• 18:05We can talk about this offline maybe,
• 18:07but it's still interesting nonetheless
• 18:09and I want to present maybe
• 18:11some some new data that suggests
• 18:13that this might not be the case.
• 18:14This is not human imaging data,
• 18:16this is a light sheet calcium data
• 18:20from zebrafish single cell data
• 18:22and it works well with what we do
• 18:24with it is something like this.
• 18:25So in addition to the the
• 18:27fluorescence traces the the fish
• 18:29are effectively moving their eyes.
• 18:31Come around during the recordings
• 18:32and we can ask this question,
• 18:34well, what do we take activity,
• 18:36something that's independent of the
• 18:38correlation structure or sorry,
• 18:39rather than that is somebody driving
• 18:42the static function productivity
• 18:44and our time varying correlations
• 18:46are each time series.
• 18:47We'll put those together into a model,
• 18:49the most derived of activity and
• 18:52connectivity and use that to try
• 18:54and predict spontaneous behavior.
• 18:56If.
• 18:56The models really only used the
• 18:59activity to explain behavior.
• 19:02Then maybe there is.
• 19:03Maybe it is true that the time
• 19:05varying fluctuations,
• 19:06the Edge time series aren't playing a
• 19:08big role here that are that they're
• 19:11really not temporally important or
• 19:14walk to anything of significance.
• 19:17We do this this is trying to
• 19:20explain fictive turns,
• 19:22and we find that we can explain about
• 19:2425% of the variance when we include both.
• 19:28I'm very connectivity and activity in
• 19:30the same model, so we're doing OK now.
• 19:33Perfect. Let me shuffle the time series.
• 19:36What we're left with is this kind
• 19:38of a garden moist war.
• 19:39Even if there's no temporal
• 19:41structure related to the behavior,
• 19:42we still get some correlation.
• 19:44It's about 10% of the variance.
• 19:47But now what happens if we shuffle activity?
• 19:50Then we're basically predicting
• 19:51with an activity alone.
• 19:53We have the same with connectivity.
• 19:54Shuffle that and leave activity intact.
• 19:57Uh-huh.
• 19:58So when we shop for activity,
• 20:00destroy the activity structure.
• 20:02We don't hurt our correlation very much.
• 20:04On the other hand,
• 20:06when we destroy the productivity,
• 20:08we get a bigger decrement.
• 20:09So suggesting that at least
• 20:11in this particular measure,
• 20:12this particular measure is
• 20:14spontaneous behavior turning.
• 20:16I think that's something that
• 20:18edges activity contributes,
• 20:20something not immediately
• 20:21accounted for by activity.
• 20:23This is only part of the story.
• 20:26There's another measure,
• 20:27this is eye movement.
• 20:28The other way around,
• 20:30if you're destroying activity,
• 20:32relation goes way, way down.
• 20:33Your model sucks basically.
• 20:36Destroy the overall structure
• 20:38in the primary connectivity.
• 20:40You don't hurt much,
• 20:41so it's really context dependent.
• 20:43Some measures are better
• 20:44predicted by activity,
• 20:45so by time varying connectivity.
• 20:48And we also using the same approach in
• 20:52pit different time varying connectivity
• 20:54measures against one another.
• 20:55And lo and behold,
• 20:56some of the better performing measures are
• 20:59the these Co fluctuated in time series,
• 21:01these Edge time series they're
• 21:03outperforming the sliding windows
• 21:05by a pretty good margin.
• 21:07OK,
• 21:07now and I promise you and circle
• 21:09back to the question we started with,
• 21:11I'm going to be trying to develop
• 21:13an edge centric
• 21:14approach for human imaging.
• 21:16Basically some way of transforming
• 21:18no data into these edge edge
• 21:21graphs and turns out we can do it.
• 21:24Basically once we have our Edge Time
• 21:26series calculate their correlation
• 21:27structure or some some measure of
• 21:29similarity and it gives us an edge by
• 21:32edge matrix fully weighted in signed,
• 21:33we can cluster it.
• 21:35When we do that we can our edge.
• 21:37Level clusters which are
• 21:39fundamentally overlapping and we
• 21:41can calculate things like the.
• 21:43Degree to which one particular region
• 21:45has multiple affiliations or not.
• 21:47So we can do all the things we set out to do,
• 21:49but also the end punch line would be a
• 21:52little weaker than that that story there.
• 21:56And I'll end with this slide.
• 21:58My claim here is not that we should
• 22:00all be doing edge centric things.
• 22:02The claim is not that you should all be
• 22:05equally code fluctuation time series,
• 22:07but rather it should be viewed as a
• 22:09complementary tool that we have already.
• 22:11Right. We're interested in brains.
• 22:13We're interested in how they're linked
• 22:15to behavior and how they how they work.
• 22:16Activations tell us something,
• 22:18connectivity or traditional
• 22:20networks tell us something else.
• 22:22And hopefully, you know,
• 22:23there's something more thoughtful.
• 22:25Not unique,
• 22:26not not accounted for by networks
• 22:28reactivations at the edges and movements.
• 22:30New approach can meet back and tell us.
• 22:33At that point I'll stop and
• 22:36open this up for questions.
• 22:44Sage bunch.
• 22:56Yes. The question is about the dependency
• 22:59presumably on parcellation INS for the
• 23:01measures that we were interested in.
• 23:02In the first papers,
• 23:03there's very little dependence and
• 23:05so far we haven't done a lot of
• 23:07reading behavior association things.
• 23:08I'm sure that there is,
• 23:09like everybody else here is
• 23:11talking about some reasonable
• 23:13dependence there and will vary,
• 23:15but we haven't seen it so far and the
• 23:16things that we really care about.
• 23:18Start with just.
• 23:25The level.
• 23:27Can you repeat the question,
• 23:29the question was then so for instance
• 23:32the zebrafish where we actually
• 23:34have like fine scale single cell
• 23:36data and we understand that stage
• 23:38was it was like. Can you just?
• 23:44Yeah. So this is you're bringing
• 23:46up one important point.
• 23:47I'll touch on this first.
• 23:49So when you build Edge time series
• 23:52or Co fluctuation time series,
• 23:54you have N parcels, voxels,
• 23:56whatever, you get that squared.
• 23:59So it becomes a real memory challenge,
• 24:01especially if you're working with boxes.
• 24:03We found some sneaky ways of
• 24:06circumventing that issue,
• 24:08holding a little bits of memory,
• 24:09doing some operations on them,
• 24:11holding something else in memory,
• 24:13doing some.
• 24:13But it becomes really challenging at the
• 24:15voxel level or even at the single cell level.
• 24:17Those are the that's not trivial to do.
• 24:19And to your question,
• 24:21we we haven't done like an exhaustive
• 24:25scale dependence with these days.
• 24:27We haven't measured that explicitly.
• 24:29I think that's an important point.
• 24:31Let's go to machine.
• 24:34So there is. Cover this.
• 24:40Spraying is very, very different.
• 24:44Yeah, that's right.
• 24:45They're tiny little.
• 24:47Couldn't talk.
• 24:49So I wonder if you kind of called
• 24:51disease reduce right up into the
• 24:52different seconds tective maybe
• 24:53look at the spot or whatever
• 24:55you know predictions change in
• 24:57these different departments have.
• 24:58It would be very easy to do.
• 25:01We have all of the atomical labels for
• 25:03for roughly where those those cells
• 25:05live and you could do this I think
• 25:07would be an interesting question.
• 25:09And I should add, we have preliminary
• 25:11data that's along the same lines
• 25:13using human movie watching data.
• 25:14It's a little more complicated.
• 25:16It's not spontaneous behavior.
• 25:17The question is can you explain something
• 25:20in responses to particular features in the
• 25:23movies using activity versus connectivity,
• 25:26putting them together,
• 25:28training explanation and
• 25:29the results again mixed,
• 25:31but there are some.
• 25:32Features where it's activity dominated,
• 25:34others where it's activity dominated Julia.
• 25:39I was wondering if they were trying to
• 25:42look at fluctuation instead of looking at
• 25:44the similarity down to looking for rather.
• 25:50And that it only relates also.
• 25:53Photography of the great, right?
• 25:54You know that oscillate at
• 25:57different different ranges.
• 26:00Like frequency dependence of this we
• 26:02have we have not this is something we
• 26:04I mean it turns out the people invent
• 26:06invented this way before us we didn't
• 26:08invent anything here it's correlations.
• 26:10There's a Christian Beckman paper
• 26:12in like 2016 where they calculate
• 26:14code fluctuations they don't,
• 26:16they don't they don't recognize or
• 26:18pursue that it's a connectivity
• 26:20over time and over that as this
• 26:22decomposition so we we haven't
• 26:23done I think it's something that's
• 26:25interesting but even our own internal
• 26:28bandwidth Ward really able to pursue.
• 26:30I think it's important question.
• 26:34There's an online question which is.
• 26:39How good is this method in
• 26:40catching moment to moment dynamics
• 26:42and what kind of parcellation
• 26:43best suits this analysis?
• 26:46How good is it?
• 26:47A catching moment to moment dynamics?
• 26:49And we end up building networks at the
• 26:52whatever frame rate you acquired your data.
• 26:55It isn't noisy, it's you're basically using
• 26:58single frames to estimate Co fluctuations.
• 27:01Again, this is not a good
• 27:03answer to your question,
• 27:04but given the bandwidth that we
• 27:06have internally for addressing,
• 27:07we're pursuing these like
• 27:09edge centric projects.
• 27:10We haven't actually looked at like
• 27:12a moment into how we haven't done a
• 27:14proper benchmarking study in that way,
• 27:16just that specific.
• 27:19Uh, attribute moment right here.
• 27:24Two questions. First, how did
• 27:26you vectorize your next series?
• 27:29For which that.
• 27:33Is there on this slide or ohh boy,
• 27:36don't go back. OK?
• 27:37There's a rule in my old land
• 27:39that every slide you go back
• 27:41you with somebody of beers.
• 27:42It's like really disconcerting.
• 27:46So I know which one you mean.
• 27:48I have to go back. Sorry, Todd.
• 27:50Sorry.
• 27:50I can find it really easily.
• 27:54Like this?
• 27:57Yeah, so each instant it's the.
• 28:00It's all pairs of nodes.
• 28:03Basically every edge in the
• 28:04network is A is a column,
• 28:07so every column here is a represents
• 28:10the upper triangle of a square,
• 28:12node by node matrix.
• 28:15So if you have 200 nodes,
• 28:1719,900 elements, yeah, yeah.
• 28:23The features you might get presented
• 28:26our submission name orthogonal to
• 28:28spiritual connectivity and activation
• 28:29will be able to pull all these things.
• 28:34$1,000,000 I mean to
• 28:34me it's a very.
• 28:37Yeah, I think this,
• 28:38this is the real question and it's
• 28:40something I think like ******* get
• 28:42yelled at when after after this.
• 28:44I know there's people in the crowd,
• 28:45but I think don't share my views
• 28:48about whether there is meaningful
• 28:50temporal information when there's
• 28:51all just static connectivity or
• 28:53activity or something like that.
• 28:55So I think we took a first step
• 28:58in that zebrafish model where you
• 29:00you put them together, right.
• 29:02You one of them has unique explanatory
• 29:04power that the other is not vice versa,
• 29:07but putting them all together.
• 29:08That incoming connectivity,
• 29:10the edge stuff into a single
• 29:12modeling is kind of challenging.
• 29:15Good ideas for how to do this, yeah.
• 29:19Yeah, so I great talk always.
• 29:24The the thing that I was that jumped
• 29:26into my mind that I I might have missed.
• 29:29So just wondering how I'm different sort
• 29:31of strategies for dealing with given
• 29:32that you know a lot of these networks
• 29:34are being specifically focused on these
• 29:36high amplitude fluctuation points and
• 29:38you're showing that that this happens
• 29:40typically like the end of movies, right.
• 29:42Wondering like how different motion
• 29:45correction strategies and regression
• 29:46people liking all these things kind
• 29:48of play into changing how these
• 29:50networks end up looking more like
• 29:52the how the algorithms play out.
• 29:54Yeah,
• 29:55we
• 29:55haven't done an exhaustive search of all
• 29:58possible motion correction strategies.
• 29:59The person to talk to
• 30:01disappear as Josh Vasquez,
• 30:02who did all of the processing for this here,
• 30:05so we can corner him after he's done
• 30:07seeing the one interesting thing I'll
• 30:09mention is that behind budget frames there,
• 30:11there's not a very clear relationship between
• 30:13motion and amplitude and the connectivity.
• 30:15But the one point that we find is a high
• 30:17amplitude frames tend to be the ones
• 30:19with the lowest instantaneous motions
• 30:21is the bars or free license placement,
• 30:23however you measure it.
• 30:24Maybe it's just like signals coming
• 30:27through at that point just knowing
• 30:29but it's that we we we see that
• 30:32relationship and whereas the lower
• 30:33the highest motion frames tend to
• 30:35be the troughs of the preserves is
• 30:37extreme right where like there's a
• 30:38lot of music music when you see high
• 30:41amplitude pretty good signal almost
• 30:43certainly not motion contaminated,
• 30:45the lowest end budget frames
• 30:48almost certainly motion contended.