Janine Bijsterbosch “Evaluating intrinsic brain organization in individuals to identify causes of network overlap”

March 07, 2023

Information

ID: 9610
To Cite: DCA Citation Guide

Download Transcript

00:07Our next speaker is Janine from Washu.
00:11Thank you. I'm excited to be here, right?
00:15This is one of my favorite conferences,
00:17but snowboarding, it's got many of
00:19my favorite people good discussions
00:21like we've already had so. Can't wait.
00:25I would class this hawk as like the
00:28sticking my head in the sand and talk.
00:31Because I agree with a lot of
00:32things that have been said so far.
00:34What I'm actually going to do is do a deep,
00:35deep dive on just one study
00:37that we've done in the lab.
00:39And I'm largely going to ignore behavior
00:41altogether and just look at some brains.
00:43So let's get back to that. Hope
00:46you're on board with that. So
00:48I'm going to talk about evaluating
00:50weighted network organization and
00:52individual differences within that
00:53and maybe what that might tell us
00:56about overlapping organization.
00:58So I don't think I need to
01:00convince anyone here that there are
01:02important individual differences
01:04in functional brain organization.
01:06These are a couple of studies using
01:09different methods from people
01:10in the room showing just that.
01:12And so these studies are showing
01:15strikingly large differences in
01:17functional organization between
01:19people that are relatively stable and
01:22therefore seems to not just be noise,
01:24but they may be meaningful.
01:29Within individuals.
01:30In some of our earlier work we did
01:33indeed show that these and this is
01:35the only behavior built in here,
01:37that these might be meaningful and
01:40that we did this was a canonical
01:42correlation analysis using human
01:44Connectome project data.
01:45We fed in a bunch of subjects
01:48specific networks and we tried to
01:51predict behavioral measures and
01:53we did indeed find that spatial
01:55network topology predicted behavior
01:57relatively well and and better most
01:59importantly than some other features.
02:01Commercial state,
02:02they don't like network agencies
02:04are going to matrices.
02:06And so this video is showing two networks,
02:08one on the left, one on the right.
02:10And kind of as it can as we cycle
02:12through the behavioral continuum
02:14of kind of more negative traits
02:16to more positive trades,
02:18how these two networks have more of in
02:21terms of their shapes as functional behavior.
02:24And to some degree this has been
02:26replicated again by some people
02:28in the room using a different
02:30approach but also showing that.
02:32And accurately capturing these
02:34individual differences in the spatial
02:36ecology of resting state organization
02:39helps improve behavior prediction.
02:44In some follow up work,
02:45we looked at the role that
02:47network overlaps plays in this.
02:49And so here there's these are two kind of,
02:52you know, cartoonish networks that might
02:54in some people not overlap at all.
02:56And then there might be a continuum of
02:59individual differences in in terms of the
03:02degree of overlapping network organization.
03:04And we found that that seems to be
03:06important with all of the caveats
03:08of behavior prediction beforehand.
03:09So I'm showing you the same video
03:11here of cycling through that
03:13same kind of continuum of.
03:14Negative to positive traits,
03:16but instead of showing individual networks,
03:19I'm showing the essentially the sums or
03:21the number of at each vertex number of
03:23networks that contribute at that vertex.
03:25So the brighter the yellow,
03:27the more overlap there is.
03:28And this actually explained most
03:30of the association between our
03:31network topology and behavior.
03:33So this seems to be something
03:35that is maybe a a feature that is.
03:42So. We the the interest so far has
03:44kind of touched on I've shown one
03:47brains because I I promised you
03:49would get back to the brains but I've
03:52shown you features from them kind of
03:54from different analysis pipelines.
03:57This is the analysis pipeline
03:58that we use a lot.
03:59It's probabilistic functional
04:00modes or Profumo or PFM.
04:02It was developed by Sam Harrison
04:04and and Steve Smith in Oxford.
04:06It's essentially like a a slightly better
04:09version of ICA push or dual aggression.
04:13So it's.
04:14On other product model, just like ICA,
04:16but it has some nice features
04:19in that it's it's hierarchical,
04:21so it has both a group model
04:24estimator in there and I'm subject
04:26specific estimator in there,
04:27and it iteratively optimizes both.
04:30It also removes the independence constraint,
04:32which is of course important if
04:34you want to accurately estimate
04:36overlap by using some other priors
04:38in this Beijing framework,
04:39like the response function.
04:41One thing I'll say here that
04:43is a little bit more.
04:44Philosophical or maybe a point for
04:46discussion later at the reception is
04:48that I think we have so many different
04:51ways of analyzing and like reducing
04:53the dimensionality of resting state data.
04:55And I've I've showed many of them here and
04:59that's that's amazing and that's useful.
05:01But at the same time I think it
05:04sometimes makes it confusing and
05:05difficult because sometimes I find
05:07it hard to understand how my result
05:09map on to results that someone else
05:11might find with the difference kind of
05:13method in a different representation.
05:15And so I think as a field we can
05:17do more coming to the kind of
05:19interoperable AI and what our values are.
05:21I think we can,
05:22it's often the same underlying source
05:24of variance in the data that might
05:26be driving our results that we're
05:28looking at from different angles.
05:29We're using different methods.
05:30And so I think to to kind of boost that,
05:33I think we can do more in terms
05:35of trying to build bridges across
05:37different analysis or different
05:39brain representations to try to
05:40kind of improve our understanding
05:41of the relationship between our
05:43way of looking at the data.
05:46And the kind of underlying
05:47source of variance in the data,
05:49that's just philosophical
05:50as I'm going to get, so.
05:53Back to grade?
05:56So what do I want to ask?
06:00The first question that I want to ask is,
06:02can I'm going to focus on human
06:04can we reliably estimate weighted
06:06resting state networks using
06:07data from only a single subject?
06:09And that might be, you know, Profumo
06:11was explicitly developed to have like,
06:13this group and subject level hierarchy
06:16to achieve correspondence across people.
06:18And so this is kind of moving
06:20away from that goal of perfumo,
06:22but I think at the same time there's many.
06:25Applications in which we want
06:26to do just this.
06:27You know, if we're doing a deep
06:29phenotyping sort of midnight scan
06:31Club type study where we're scanning
06:33few people there very long time,
06:36we want to estimate networks
06:37just from individual people.
06:39If we want to do some translational
06:41work in animals,
06:41which I would say is something that
06:43we have to do more of to really gain a
06:46better understanding of what we're measuring,
06:47then you know,
06:48we're doing monkey work.
06:49We can't get that many monkeys that
06:51we might want to do individual
06:53exclamation from the single subject.
06:55And also ultimately if we want to
06:57translate this to patients and we're
06:59seeing that individual 65 year old
07:02that's even was talking about in
07:04front of us and we want to make a
07:06prediction on that page and we'll
07:07presumably we want to just be able to
07:09analyze the data from that patient.
07:11I mean we'll probably compare it in
07:13some sort of normative way to other people,
07:15but we don't have to have the computational
07:18cost that we analyze everything.
07:19So I think this is an important question
07:22from a few different applications.
07:24Then I want to ask is is spatial
07:26overlap present in individual subjects?
07:28It could just be driven by the group bias
07:30because group averaging is essentially
07:32smoothing and blah blah tying our things.
07:35So maybe it's just a groupies aspect.
07:37And then the last thing I want to
07:40ask or touch on a little bit is,
07:42is is try and get to the mechanistic
07:44level and try and ask can we understand
07:47better what gives rise and how should
07:49we understand these features of spatial
07:51overlapping that work organization?
07:55So I use data from the
07:57Human Connectome project.
07:57This is only 20 subjects because
07:59these are the all of the 20
08:02subjects that I have 12 scans.
08:03So these are people that did 3T3T3 tests
08:07and and 78 includes 78 twin pairs.
08:11And we did perfume on a
08:12bunch of different ways.
08:13So a standard group analysis,
08:15a single subject analysis,
08:17a splitting half single
08:19subject analysis and kind of
08:21systematically adding more data in.
08:23Just look at data requirements analysis.
08:26Um, A group analysis about
08:28matching the amount of data that
08:30we have in individual subjects.
08:33And then we did this at
08:34the dimensionality 20.
08:35It's always arbitrary,
08:36but we wanted to look at kind of
08:39the canonical large scale networks.
08:40And then within these 20 networks,
08:42we look for networks that have
08:44both high within subject split
08:46reliability and also at a high kind
08:49of subject to group correlation.
08:51So both testing kind of
08:53reproducibility at the subject
08:54level and correspondence with the
08:57group level or across subjects.
09:01Proposed brains here, a lot of brains.
09:04These are the group modes.
09:05So the top row is using the the main
09:09standard analysis in the bottom row is
09:11essentially the same thing with using less
09:14data to match the amount of data that
09:16we have in the subject level results.
09:18I think there's no big surprises here.
09:20I think many of you probably
09:22recognize a lot of these networks.
09:24You get less SNR when you use less data.
09:27No big surprise there either.
09:30That seems to impact the negative.
09:31It's little bit more so
09:32than the positive weights,
09:34which might be interesting.
09:35But let's get to the individual differences.
09:38So this here is the same thing
09:41on the same set of results,
09:44but from an individual subject.
09:45So the top row this time is when the analysis
09:48is run using data from just that subject.
09:51So there's no group kind of information
09:54that the model sees at any one point.
09:56And the bottom row is the same
09:58subject specific estimate from the
10:00hierarchical model in which there is
10:02kind of the group level above it.
10:04And so it looks,
10:06you know,
10:07there's a lot less lobbing this as I was
10:09saying compared with the previous slide,
10:11you know,
10:11there's a lot more kind of detail in there.
10:14If you zoom in on it a little bit
10:16in some areas there's you seem to be
10:18getting that really kind of subject
10:20specific organization and those features.
10:22And if I go to a different subject and
10:25kind of toggle back and forth between them,
10:28then you can see,
10:28you can start to see,
10:30you know these are recognizably the
10:32same networks and yet you see these.
10:34Relatively substantial shifts in,
10:36in and and kind of changes in organization.
10:44This is a sort of winner.
10:46Take whole representation or find the
10:49biggest representation with for all subjects.
10:51I'm not going to and they're kind of
10:54grouped into twins for the sake of time.
10:57I'm not going to spend too much time on this.
10:58You can stare at this for a long
11:00time and I don't know that,
11:02but I want to talk about this.
11:03So these are some results.
11:05On the most left hand side
11:07is test retest reliability.
11:08So within individual subjects
11:12there's a dot is a is one of the 12.
11:14That works.
11:15And is there for all of the subjects.
11:17I highlighted in red and green,
11:18the example subjects,
11:19just so that I can show you that they're not.
11:21They're like middle of the road subjects.
11:23They're not particularly fancy
11:24or good or bad.
11:26The 2nd row in is subject
11:29to group similarity,
11:31so that's essentially quantifying these two.
11:34The the correlation between the
11:36top and the bottom row here.
11:38And that's interesting.
11:39One of the things I think we worry
11:42about when we're estimating from
11:44suggesting single subject is correspondence.
11:46You know how do we,
11:48how do we relate it back to other
11:49people and can we compare apples
11:51to apples or are we now you
11:52know comparing apples oranges.
11:54And I think the results were kind
11:56of surprising to me in terms of at
11:58least at this level of dimensionality
12:00which is a relatively low you know
12:02large scale network dimensionality we
12:05we get pretty high correspondence.
12:08Without enforcing that in in a
12:11in in the model itself.
12:13The other interesting thing is that
12:14you kind of see this stepping effect.
12:16So these two on the left,
12:17these two columns on the left
12:19are kind of within subjects,
12:21the next two are twin and the next two
12:23are between subject for non twins.
12:26And so you can see you know within
12:28subjects is most reliable twins or
12:29next most reliable and then you
12:31get a a slight decrease to decrease
12:33in the next one.
12:34Another thing you see is this
12:37systematic kind of bomb.
12:39So this this twin similarity,
12:41the middle one here is using
12:43individual data and then the next
12:45one I was using group informs data
12:48and then it's the same thing here
12:50for between non trained subjects.
12:52Individual day runs through runs
12:54and so you do see that group by
12:57aspects there a little bit,
12:58but you see that you achieve correspondence
13:01even without needing that group bias.
13:05I'm not going to dwell on this.
13:07You the more data you get,
13:08the better you do.
13:09And obviously you're going to need
13:11high quality data to do this.
13:15The next question was is there overlap,
13:18is there a spatial overlap?
13:19And the answer is yes there is.
13:21So on the left hand side here is a
13:24is a kind of overlap matrix at the
13:26group level and on the right hand side
13:29it's showing a surface area for each
13:31individual of how many networks overlap
13:34within that subject within that area.
13:37And this is showing that
13:38on the on the spatial map,
13:40so you can see again for the
13:42same two example subjects,
13:43you can see that the main areas of
13:46overlap are in that temple parietal
13:49occipital junction broad area,
13:52which I think is interesting
13:53because I think with other methods
13:55like perturbation methods that's
13:56also where you see a large amount
13:58of individual differences.
13:59So that again comes back to this
14:01question of how do we how do we interpret
14:03results from different methods and
14:05how do they kind of tap into the same.
14:07Underlying variance in the data.
14:12And so with my last few minutes,
14:13I want to I want to try and understand.
14:16I think these overlap areas are
14:18are interesting and important and
14:20so I want to try and understand
14:22them a little bit better.
14:23I have 3 hypothesis about
14:24what might be happening.
14:25If you have more please let me know.
14:29The first one is that it the, the, the,
14:31the kind of mixing I've also so it might
14:33be this is the idea of interdigitation.
14:36So maybe there's like ribbons of
14:38network one and network two kind of
14:40you know alternating and we're not
14:42quite resolving that based on our data.
14:45The second hypothesis is a is a
14:47kind of a dynamic hypothesis.
14:49So maybe what's happening in this
14:50overlap area is that it's either part
14:52of network one or part of Network 2,
14:54but it's kind of going back and
14:56forth in time between them.
14:57And of course given that perfumo
14:59is a is a stationary method that
15:02would show up as as kind of overlap.
15:05Or maybe the third one and maybe the most
15:08exciting one is a coupling hypothesis.
15:10Is this area really kind of part of
15:12both networks at the same time and
15:14that could be really cool in terms of,
15:16you know,
15:16maybe there's some sort of integration or
15:18meaningful processing and cross network
15:20processing happening in these regions?
15:24And so I'm going to try and ask this
15:28question, trying to test these different
15:30hypothesis using two networks of interest.
15:33So there's the red one, there's yellow one,
15:35you can see there's the areas of orange
15:38are kind of overlap between them.
15:40And if we look at these are
15:42the two examples subjects.
15:43So you see there's individual differences,
15:45some have more overlaps, less overlap.
15:48What I want to do is essentially isolate
15:51vertices that are just network one,
15:54nothing else, just network 2, nothing else,
15:57and just overlap between those two and
16:00they're kind of third or fourth networks.
16:02And so I'm kind of going to clean this out.
16:05There are some bird seeds that are also
16:06going to have contributions from some
16:08of the other networks that we estimated.
16:10Just want to keep this like as clean
16:12and control the comparison as possible.
16:14So that's what I've done here.
16:15So red vertices are in the.
16:19Overlap between these two specific networks
16:22and black R1 and green or the other.
16:25And so now you can estimate average
16:28time courses across these vertices,
16:30and we get a a time course for
16:32each of these three things.
16:34And we can do that.
16:35So this is them completely separately
16:37for each of the 12 people.
16:39Rewrite a few more completely
16:40independently and do all of these steps.
16:43And so now I'm going to we have
16:45the true overlap time series,
16:46but we can also create hypothesis
16:49based overlap similarly simulated
16:50versions of the overlaps comma series.
16:53So for the mixing hypothesis one of
16:55the things I can do is for every for
16:57half of the red vertices I can just
16:59pick a random black vertex and for the
17:02other half I can literally rather than
17:04green vertex and then I can re average them.
17:06And so I get a simulated version using the
17:08data from this individual but not using
17:10any of the red parts of the data, right?
17:13I can do the same thing for the
17:16coupling hypothesis,
17:17so at any time point I can just sum up
17:19the black and the and the green one.
17:21Or maybe multiply them together.
17:22Linear and nonlinear kind of coupling
17:27hypothesis?
17:28Simulated versions?
17:28And then for the switching one,
17:31I can kind of just go back and
17:33forth between them in multiple ways,
17:34different different orders,
17:36different time Windows,
17:38different ways of doing that.
17:39And so now I've come up with
17:42essentially 8 different simulated
17:44versions of the overlap time series
17:46to test these different hypothesis.
17:49And so the first thing I can now.
17:52Ask is which of these countries
17:54are most similar to the actual
17:57original true overlap time series.
17:59And so there's some interesting patterns
18:02here and we talk about that a lot more.
18:04But the winner visually and statistically
18:07is the linear additive coupling.
18:10So that is coupling hypothesis
18:12that just sums up to other regions.
18:17Without having time to go to a
18:19lot of different details, we can.
18:21We also did hidden Markov modeling to get
18:24dynamic states from the three time series.
18:27So network one time series,
18:29network two countries,
18:30and overlap time series.
18:32You can do that using the original
18:34overlap time series and using
18:36all of the simulated versions.
18:38It's a little bit noisier and that
18:39makes sense because it's more
18:41complicated model that we're fitting,
18:42but it's nice to see that the same
18:45kind of hypothesis is winning here.
18:48Which is the linear additive coupling 1.
18:51And so with that,
18:52my first question is can we reliably
18:56estimate these weighted networks?
18:58And the answer is yes,
19:00given sufficient high quality data.
19:03This spatial overlap present, yes,
19:05we see we see significant and pretty
19:07reliable and reproducible spatial
19:09overlap at the individual level.
19:11And then the third question was,
19:12well,
19:13what mechanisms might give rise to that
19:15at this at MRI resolution given the
19:18constraints of what we've done here,
19:21it appears to be the linear
19:23additive problem hypothesis.
19:24I have a lot more questions
19:26about that because of voxel ISM,
19:27big space like we have a lot of
19:29neurons in a voxel and so there might
19:31well be interdigitation at the level.
19:33Smaller than a voxel,
19:35a TR,
19:36even with multiband is also still pretty
19:38long compared with neurons firing.
19:40And so there might well be dynamic
19:43switching at the level at a temporal
19:45resolution below beyond them IR MRI.
19:48And so I'm I have some funding to do
19:50some animal work to try and get more
19:52closer to that and kind of get beyond
19:54the level of MRI to ask those questions.
19:59So with that,
20:01I will thank my collaborators for this
20:04particular project for at Oxford and Washu,
20:07my lab,
20:08my funding and the datasets
20:10that we all work with heavily.
20:13And thank you very much for your time.
20:24Really cool stuff. I would refer
20:28to up your thoughtful hypotheses.
20:30What would be happening?
20:31I thought of two more potential.
20:36Could it be that your dimensionality is
20:40not high enough and there are certain
20:44additional networks in there that are,
20:46I would say in of that,
20:48I think you could either hypothesize
20:51a network kind of in between in
20:54network space like you're, you know,
20:56you have like a default thing and there's
20:58something in between that, right?
21:00It's like halfway one and
21:01halfway to the other, right?
21:03And then when you have a lower.
21:04Really it kind of they overlap, right?
21:08And then the other one I was thinking
21:10of is if you had a network that is not
21:13asserting you're going to mentality,
21:15it's kind of neither one,
21:17but is spatially in between, right.
21:21And you're on especially thinking of
21:24this temporal parietal occipital area
21:26where there's stuff going on with
21:28there and I expect we haven't found
21:30what the actual representation that
21:31representation is in there and if there's,
21:34if there's.
21:36If your signal in there,
21:37but it's not very well captured by any
21:39of the sort of well known networks,
21:42could they just be kind of assigned to
21:45to whatever to to the 20 dimensionalities?
21:48Kind of halfway to the halfway?
21:50OK, so so the question, well,
21:52there was a suggestion of
21:54two alternative hypotheses.
21:56Here. Right, right.
21:58The first alternative hypothesis being
22:01maybe the dimensionality is too low,
22:03and the second hypothesis being
22:05maybe we just don't know what to do
22:07or know none of our models really
22:09knows what to do with those areas.
22:12Yeah. Yeah.
22:13And so we have not in this specific work,
22:17but in other work we have pushed
22:19up the dimensionality of of perfumo
22:21in particular because it behaves
22:23a little bit different from ICA.
22:25For example,
22:25when you push up the dimensionality
22:27and you see this kind of,
22:28you do see this you,
22:30you continue to get the canonical
22:32big picture networks in Profumo even
22:34if you push up the dimensionality
22:36because they don't have to
22:38splinter down smaller because
22:39of the constraints and priors.
22:41You do that to see additional networks
22:43but I I don't think there's networks
22:46map on particularly well they're like
22:48it gets like the whole network and
22:50then it gets sub networks within that.
22:52But I don't think that maps on
22:55particularly closely on to this.
22:57But I I I can move closer into that.
22:59I think you are right that.
23:02They're just there seems to be a lot
23:05of complexity in that temporoparietal
23:07acceptable kind of area and depending
23:09on what model you throw at it,
23:11it's going to show up differently.
23:13And I think we've seen that
23:14in in some of your work.
23:16And I think that there's a lot of and
23:18that's what I that's why I was talking about.
23:20I feel like we need to do better
23:22in connecting our results and like
23:23seeing you know what results with
23:25different methods map on to each
23:26other because I think that we there
23:28might be some insights in there.
23:30We do more of that.
23:31But yeah I agree that.
23:33There's something really interesting
23:34going on in this area and and,
23:37and I'm not saying that this is
23:38the final answer to it, I think.
23:41Simon.
23:43The interesting thing about that
23:45area is that it's also one of the
23:47most variable areas starting already
23:49from macro anatomy. So the question
23:51is how much in terms of also trust
23:55registration is mismatch may play into them.
23:58It's it's it's always that area
24:00where it seems smarter anatomy,
24:02but rather me sunshine.
24:04Maybe it's just too variable to
24:08really make good statements.
24:10Yeah, I think that kind of so the
24:13question was is that error that
24:14area is also hard to register
24:16and so is that is that is that
24:18fundamentally driving into this.
24:20I think that actually gets back to
24:22your point about is it a compound
24:24or is it something meaningful
24:26because one of the things we did
24:27in in the one of the earlier papers
24:29where I had the video of it.
24:31We did the the we we drove the CCA with
24:34behavior from the warp fields as well.
24:37So the stuff that we throw away usually
24:39and that predicts behavior really well.
24:44And so you're right,
24:45like we're not doing it probably
24:47maybe as good a job registering,
24:48but but maybe once we do,
24:50we throw away the individual
24:51differences that we care about.
24:53And so I don't know the answer to that,
24:54but that's just that with that out there.
24:58And so I guess I have that.
25:06MSN.
25:09Oh yeah, the data are MSM all registered,
25:12so the we did the best we could.
25:16Also like to get the overlapping individuals.
25:21Level from the type of the group to select.
25:25So I think my question is that you
25:27specifically pick the ones that have
25:3017 and like all 12 sessions that you
25:32didn't just like you just want more
25:35sessions or you actually compare that?
25:37Do they have the same kind
25:40of overlap networking?
25:44I'm guessing they also have for the
25:49individual overlaps of consistent
25:51across different sessions must.
25:56Yeah, so the question was,
25:57did we do anything beyond just using
26:00the extra data with the 7T data?
26:03We didn't in this,
26:04we just used it this extra data, but I
26:07think those are interesting ideas to see.
26:10There didn't seem to be any.
26:12Systematic differences with the
26:1417 data with the other data.
26:17Let's go, Ruby.
26:31Yes. So the question is the overlap
26:34kind of relies on the thresholding and
26:36there's different ways you can do it.
26:38So we if if you just look at
26:40the spatial correlation matrix,
26:42so that's just like straight up correlating
26:44with a few the within subjects,
26:46the weighted network.
26:47So that doesn't require any thresholding.
26:52And so we did that for some of the stuff,
26:55yes, if you want to kind of for for the
26:57simulated versions and stuff like that,
26:59you do need to put a threshold.
27:01We put a fairly stringent threshold so that.
27:05Areas had to have yeah, so.
27:09I don't think it changes much with
27:11thresholding and particularly with.
27:13The nice thing about Profumo is that it has
27:15kind of like a binomial distribution with
27:18whether it's sampling a signal or noise.
27:21So it has a relative,
27:22but even though it's weighted,
27:23it doesn't have a lot of kind of
27:25weights in the middle because
27:28of the because of the model.
27:31Period. Are you?
27:38Anatomical problem.
27:43For example. Like more
27:46today?
27:50Will you repeat that?
27:51Yes, the question was have we looked
27:54at the structural connectome and and
27:56how this relates that we are currently
27:59doing a project in the lab trying to
28:01link perfumo networks and individual
28:03differences to to structural networks.
28:05I don't have any big answers from that.
28:08But yeah, I think it is
28:10important area for sure.
28:11We are also planning to do some of this
28:14in monkeys afterwards which I think will
28:17be just monkey MRI data to start with,
28:19which I think would be interesting.
28:20To get at some of the announcement
28:23and related questions.
28:26When the last question and then
28:28I'll make a comment. Go ahead.
28:33Possible. I wondered how you were.
28:42Like the numbers one is
28:45sort of. You can also.
28:49Those Peppers are not
28:51associated with support.
28:56I noticed that you did some post.
29:02Yeah. And so the question is what
29:06are these 20 networks really,
29:08which we did some testing at
29:11different dimensionalities I think
29:13between like 10 and 30 years,
29:1540 to see and and actually with Profumo
29:19you get pretty similar networks if
29:21you go higher because it doesn't,
29:23it doesn't have to splinter down
29:24smaller because it doesn't have
29:26that independence constraint.
29:27So the networks,
29:28these 12 networks are things that
29:29you see that are relatively what
29:31I would say Canonical that you
29:33see with relative reliability.
29:35And they do to some degree map on to so.
29:38So the, the mapping,
29:39the mailing that we did was
29:41based on the wooden paper, the.
29:44We didn't go at the I I told you
29:46this was a sticking my head in the
29:48sand and talked and I was looking
29:49to talk about behaviors of the 20
29:51subject to talk about behavior right.
29:54But it is the case that so from
29:57that video at the start I showed you
29:59like the default mode like and the.
30:02The cognitive networks are the ones
30:03that are the most important for
30:05behavior and that also contribute
30:07a lot to that overlap in that.
30:14Great. Thank you very much everyone.
30:16Comment which you already emphasized
30:19or you already mentioned,
30:20but I want to emphasize which is this
30:22potential for complexity of the Walker level.
30:25And apparently in animal
30:26studies you look at that,
30:27it seems to me that all three of these
30:30hypothesis could be true that the,
30:31you know within a voxel because
30:34there's hundreds of thousands of
30:35neurons in there and they presumably
30:37don't all do the same thing so.
30:39Yeah, that's great,
30:41but I'm glad you got to look at that.
30:45Yeah, it's really hard.
30:48Uh, it's a good question, but uh,
30:50yeah, tough on the answer.
30:52Thank you very much, Janine.
30:53Thanks, everybody. We're.