Alan Anticevic, PhD. February 2022

Transcript

• 00:17Welcome, everyone.
• 00:18It's a pleasure today to
• 00:21have Alan antiseptic here.
• 00:23Alan, I think many of you know,
• 00:24is an associate professor in
• 00:26the Department of Psychiatry.
• 00:29As well as many other hats.
• 00:31And his was a really a pioneer in our
• 00:34in our community and bringing high
• 00:37resolution human Connectome Project
• 00:39style imaging to the community and
• 00:42then applying innovative computational
• 00:45strategies to analyzing these data.
• 00:48And then in more recent years,
• 00:49it started to do some some really
• 00:52exciting work in looking at the
• 00:55brain effects first of ketamine
• 00:56in his earlier work and then
• 00:59more recently of psilocybin,
• 01:00LSD and other psychedelics,
• 01:02which is the focus today.
• 01:04So Alan,
• 01:04thank you so much for being
• 01:05with us and we look forward
• 01:07to to learning from you.
• 01:10Thanks, Chris. Mark Gustavo.
• 01:13Anytime at all.
• 01:15It's really with this pandemic,
• 01:17I feel like I haven't seen you
• 01:20guys in forever and it's feeling
• 01:22more disconnected than ever,
• 01:24but it's it's it's really.
• 01:26I'm excited to tell you
• 01:28what we've been up to so.
• 01:30The the title is so let's see mapping
• 01:35or behavioral heterogeneity of
• 01:37psychedelic neurology and humans.
• 01:38So that's. An overly ambitious,
• 01:41probably not true title.
• 01:42So we're not quite there
• 01:44and this is aspirational.
• 01:45And So what I'd like to kind of
• 01:48talk about is as as Chris noted,
• 01:50what are the techniques and approaches
• 01:52we're bringing to bear towards
• 01:53this goal and some of the work
• 01:55that's been coming out of the lab.
• 01:57So it will be a hybrid of sort of the, the,
• 02:00the theme and the ethos of what we're doing,
• 02:03how some of the efforts in in mapping
• 02:08variation in clinical populations.
• 02:11Can be related to neuroimaging
• 02:14effects of pharmacological compounds
• 02:16with the focus on psychedelics.
• 02:18So that's kind of the idea,
• 02:19right.
• 02:19And so I I want to just disclose
• 02:23that I'm a member of the TAB
• 02:26Technology Advisory Board for Nemora
• 02:30Therapeutic Therapeutics I consult.
• 02:32For Gilgamesh and I just Co founded
• 02:35a biotech with my colleague John
• 02:38Murray that's called Manifest
• 02:41Technologies and I I say this
• 02:43really proudly because without the
• 02:46support of Yale the spin out what it
• 02:48would not be possible and so we're
• 02:50we're really excited about this
• 02:52work as it has this potential so.
• 02:54Basically the way I kind of
• 02:56think about the entire space.
• 03:01Neuropsychiatric mapping is
• 03:02really about the challenge and the
• 03:04opportunity in front of us, right?
• 03:06So there's two ways to think about it.
• 03:08And why has this opportunity not
• 03:11been realized in the field of brain
• 03:13behavioral health and I'll use this term
• 03:16brain behavioral health because I I I
• 03:19actually think it's important that we
• 03:22destigmatize this terminology, right.
• 03:25So it's a difficulties in regulating
• 03:29brain behavioral relationships that that
• 03:32we're after and then I'll talk about a
• 03:35framework for this quantitative and and.
• 03:38The neurobiological framework for mapping
• 03:40brain behavioral relationships with the
• 03:42assistance obviously of pharmacological or
• 03:44imaging as a key tool in order to do this,
• 03:47right. So what's the challenge?
• 03:50And so this is one of many challenges,
• 03:52but in my mind a very important one,
• 03:54right, which is heterogeneity.
• 03:56And by heterogeneity,
• 03:57I mean both within a human being over
• 04:00time and across people in relation
• 04:03to brain behavioral variation, right.
• 04:06And so this is an old problem.
• 04:09We know this, right?
• 04:10And so, but there are.
• 04:12Questions that arise because of this problem
• 04:16and the opportunity in front of us is if we.
• 04:19Have a drug like any compound.
• 04:23And psychedelics are a great example,
• 04:26right?
• 04:26How do we select the optimal person
• 04:29who will benefit from that compound?
• 04:31We don't have a principled quantitative,
• 04:33rationally guided framework for this,
• 04:36for anything in our field.
• 04:38And then in the future,
• 04:40right,
• 04:40if we do have a molecule right
• 04:42before we do not have a molecule,
• 04:43if we don't have a compound that
• 04:45crosses the blood brain barrier
• 04:47safely or any therapeutic,
• 04:48how do we develop one with
• 04:50individual precision as the goal,
• 04:52not patients versus controls,
• 04:55but individual people, right.
• 04:57So that's the opportunity as I see it,
• 05:00right and so.
• 05:01It's that how can we target
• 05:03specific people with?
• 05:05Quantitative precision.
• 05:06So this problem I see is mapping one
• 05:10to many levels of analysis, right?
• 05:12This is really what stands in front of us,
• 05:14so I don't have to tell you guys that.
• 05:17Polygenic disturbances and variants.
• 05:22Variations with rare mutations, right?
• 05:27Basically those are the two lowest
• 05:29level possibilities that our field
• 05:31is studying and how they can affect.
• 05:34Molecules, synapses and cells and the
• 05:36balance between those cells, right?
• 05:38That's that's at the very
• 05:40baseline of the problem, right?
• 05:42In turn,
• 05:43how do we take that information and
• 05:45map it onto system level observation?
• 05:48Some people would say this is an ill
• 05:49posed problem because it's impossible.
• 05:51There's just too many mappings.
• 05:52But you know,
• 05:53I'll leave that for debate later.
• 05:55And then finally,
• 05:56how do we link this to Spectra of
• 05:59behavioral disturbances, right.
• 06:00And so I'd like to argue that this
• 06:02mapping is fundamentally unknown.
• 06:04We don't know it,
• 06:05and if somebody argues that we do,
• 06:06I think that they're flying.
• 06:09Even in circuits where we understand
• 06:11our biology very well, like fear,
• 06:13we still can't treat PTSD, right?
• 06:15So just this mapping is not accomplished,
• 06:18right?
• 06:19And I actually think that system
• 06:21level observations at the
• 06:22level of neural systems is where the
• 06:24right link to behavior should be,
• 06:26not at the level of a synapse,
• 06:28because the heterogeneity just explodes,
• 06:30the combinatorics become impossible to
• 06:32intractable to quantify or deal with.
• 06:35So why have we not solved this problem right?
• 06:39Why is this opportunity not been realized?
• 06:41And so I think that.
• 06:44I might argue my many reasons,
• 06:46but there's some legacy barriers,
• 06:48right, that we're still
• 06:49trying to overcome as a field.
• 06:50And I and the legacy approach are called
• 06:54legacy because it's historically important
• 06:57to acknowledge that this this is.
• 07:00The framework that we have been
• 07:02operating under is has tremendous
• 07:03utility for what it was designed
• 07:05for and by that I mean DSM, right?
• 07:07It does what it was built for to do,
• 07:11which is it reliably gets me and
• 07:14everybody else to agree that some person
• 07:18has X out of P symptoms over T time.
• 07:22That it does quantitatively accomplishes
• 07:24that, so we can reliably agree.
• 07:27To categorize a human being as
• 07:29you showed X symptoms out of,
• 07:32you know, some rubric overtime.
• 07:35And then we give you a label.
• 07:36I'd like to argue that's ill
• 07:38fitting for what we need,
• 07:39actually not wrong for what it was
• 07:42built to do, just not what we need.
• 07:45And then further,
• 07:46we lack data and methods to
• 07:47quantitatively molecular benchmark
• 07:49brain behavior relationships, right?
• 07:51So the legacy approach doesn't have it.
• 07:52And then we don't have technological
• 07:54solutions to actually scale this with
• 07:56the kinds of observations that are so
• 07:58important in the area of psychedelic
• 08:00medicine for precision therapeutics,
• 08:02right.
• 08:02So this is a dystopian vision of the future,
• 08:07right,
• 08:07that I'd like to show sometimes like
• 08:10basically where we use computer assisted
• 08:12decision making or machine learning informed.
• 08:15Decisions with multimodal data where
• 08:16the brain is in the middle, right?
• 08:18I'm not taking the person out of this.
• 08:20I'm not, you know,
• 08:21reductionist to that point,
• 08:23but the organ has to be in
• 08:25the middle in my mind.
• 08:27And then we're hopefully optimizing these
• 08:29two decisions through an iterative cycle,
• 08:32right?
• 08:32But we're not there and and so,
• 08:34So what do we do about that?
• 08:36So my group here and our other collaborators
• 08:40here at Yale are really approaching this in,
• 08:43in the following way.
• 08:45So let me walk you through the framework.
• 08:48So we have to 1st agree that the
• 08:52neurobehavioral mapping problem is
• 08:54quantitatively A must, we have to do that.
• 08:57Correctly.
• 08:57And I'll explain what I mean by that, right?
• 08:59Because if if we are using an
• 09:02ill fitting framework to map
• 09:04onto neurobehavioral variation,
• 09:05then just won't work.
• 09:07We won't even translate what we want.
• 09:10Gene expression alterations right
• 09:12and disturbances in the way
• 09:14that the circuits are formed are
• 09:17it's useful information.
• 09:19We harness that from say the island
• 09:22human brain Atlas to inform our our
• 09:24our models that then can simulate
• 09:27pharmacological fMRI effects which in
• 09:29turn can be fixed individual people.
• 09:31So this is just one way
• 09:33that we're approaching this
• 09:35problem in the area of just mental
• 09:37health and and psychedelic pharmacology.
• 09:39Specifically, but there's there's many
• 09:41other tools that we're leaving on the table,
• 09:44but this is what I'm going to talk about,
• 09:46right. So again I'd like to argue
• 09:47that you know Chris asked me to talk
• 09:50about neuroimaging today specifically
• 09:51and to really have that focus.
• 09:53So I'd like to argue that new imaging
• 09:56is not an option anymore in our work.
• 09:59And by that I mean normalizing broadly,
• 10:01right? But a necessity, because here,
• 10:04here's a choice, right?
• 10:05This is literally what we had on the table,
• 10:08right? But we need this.
• 10:12And so I'd like to argue that if if you
• 10:15present the problem this way, right,
• 10:16like how can we not leverage brain data
• 10:19in the service of decision making as a must,
• 10:22specifically for things as complicated
• 10:24as psychedelic neurobiology at
• 10:26the individual level, right.
• 10:28So.
• 10:28So, OK, so how can we exploit imaging
• 10:31in the service of this goal?
• 10:33So that's the rest of the talk.
• 10:35So for those of you who haven't
• 10:38really thought about.
• 10:40Structural functional
• 10:41multimodal imaging recently,
• 10:43this is a very useful reminder of the
• 10:45resolution of what imaging straddles,
• 10:47right.
• 10:47So on the Y axis is the size of
• 10:50the observation and time is on the
• 10:54the X axis and and you'll notice
• 10:56that human imaging across modality
• 10:58straddles a good bit of the space.
• 11:00We're not helpless right.
• 11:01Like we can actually measure signals
• 11:04in various ways in relation to various
• 11:06phenomena, but we still are, you know,
• 11:08out of reach of certain levels of analysis.
• 11:10Like with human imaging with symptoms
• 11:12improving but still not quite there.
• 11:14And we also don't have just one
• 11:16way to measure this,
• 11:17right.
• 11:17We have multiple modalities that now can
• 11:20combine that give you different slices
• 11:23of the signal that this incredibly
• 11:26complex piece of tissue produces,
• 11:28right.
• 11:29And so,
• 11:29so I'm going to talk about specifically
• 11:32throughout the rest of talk about bold F MRI,
• 11:34right, which is a measure that is is.
• 11:39One of the expertise areas in my group,
• 11:41right so.
• 11:44Specifically,
• 11:44the phenomena that for the rest
• 11:46of the the talk will focus on is
• 11:49this idea of resting state, right?
• 11:51And so this is now a household name.
• 11:53I, you know,
• 11:54I don't have to go over this
• 11:55in a lot of detail,
• 11:56but I'd like to kind of just
• 11:58review the history of it.
• 11:59I always like to do this because I
• 12:01like to remind people that in 1995,
• 12:03right, broad Biswal.
• 12:05Actually observe this simply
• 12:08by taking signal.
• 12:10Out of 1 hemisphere of the motor
• 12:13cortex and asking where in the
• 12:15brain is there time dependent
• 12:17covariation of signal within a person
• 12:20and then average across people.
• 12:21And he saw this map right which is
• 12:24bilateral motor cortex comes out.
• 12:26And people thought this was junk.
• 12:28This was a compound, right?
• 12:30And they ignored it.
• 12:31In fact, he was attacked for it quite a lot,
• 12:33right? And then only a
• 12:36decade later with some.
• 12:40Advances from Michael Greicius
• 12:41and then Marcus Raichle.
• 12:43Did this phenomena really
• 12:45become a mainstream?
• 12:46And so now if you Fast forward
• 12:48what is now a decade ago,
• 12:49which is hard to believe right,
• 12:50that Thomas Yao and Randy Buckner
• 12:54actually mapped comprehensively right.
• 12:56Large scale networks
• 12:58across individuals right.
• 12:59And this is this is now a joke
• 13:03like we we know we can do this now
• 13:04and every single person right.
• 13:05But it was controversial
• 13:07then and so then 2016 onward.
• 13:09Human connectome produced sparse
• 13:11relation which gives you an index of
• 13:14the boundaries between the areas that
• 13:16is comprehensive but not definitive.
• 13:18This will probably improve it or actively,
• 13:20but I'd like to tell you like we've
• 13:22made some serious progress, right,
• 13:23with human or imaging and we can exploit it,
• 13:26right?
• 13:27And furthermore,
• 13:28this is something I will repeat
• 13:31in every talk I give.
• 13:34So the reason why the Human
• 13:36Connectome project pipelines and the
• 13:38entire effort towards so impactful,
• 13:40which is why I obviously drink the
• 13:42kool-aid because I trained there, but.
• 13:46The human brain and the cortical
• 13:49mantle is A2 dimensional surface
• 13:51wrapped around white matter that is
• 13:54about 4 millimeters thick and it's the
• 13:57size of a pizza and that geometry matters.
• 13:59It matters fundamentally when you're
• 14:01going to do precision medicine
• 14:03analytics with single subject
• 14:05human cortical surfaces, right?
• 14:07In fact,
• 14:07just before this talk I got off
• 14:09the call with one of my students,
• 14:10Amber Howell,
• 14:11deeply debating the importance of
• 14:13importance of social curvature
• 14:14and depth on a single subject
• 14:16level in relation to.
• 14:17Diffusivity of white matter tracts,
• 14:19and turns out it matters.
• 14:20It matters a lot anyway.
• 14:23Analyzing your data on the surface,
• 14:25I think is is. I must like you.
• 14:29You are blind without that to
• 14:31individual variation, right?
• 14:33So.
• 14:34When you do this right,
• 14:36then you can produce metrics in the
• 14:39geometry of the cortical surface
• 14:41that can quantify some signal in
• 14:44every cortical parcel or area.
• 14:46I'll use the term parcel because
• 14:48that's the formal definition.
• 14:49And then you can operate with those
• 14:51metrics to analyze it within a
• 14:53subject across people in various ways,
• 14:55right.
• 14:55So furthermore,
• 14:56what we've done in my group out of necessity,
• 14:59right,
• 15:00we started to use the cortical parcellation
• 15:02from the Human Connectome Project,
• 15:04which is shown here.
• 15:06These little borders,
• 15:07but then we use the network partitions from
• 15:09from other groups and they were great,
• 15:12they they worked really well.
• 15:13But then we realized subcortical
• 15:15coverage is not there.
• 15:17Like thalamus wasn't covered appropriately,
• 15:19brainstem wasn't covered.
• 15:20And you guys know that psychiatric
• 15:23medication works on subcortical
• 15:25circuits and precision of isolating
• 15:28specific voxels in relation to
• 15:30networks and parcels is really important.
• 15:32So what I'm showing you here is
• 15:34work produced by my student Lisa.
• 15:36Who's extended the cortical partition
• 15:39into networks Cortically first,
• 15:41and then studied how subcortical
• 15:44voxels covary with those networks,
• 15:47assigning every single voxel in
• 15:49subcortical brain matter to a
• 15:52network until it reached split half?
• 15:55Stability across the entire HCP sample.
• 15:58And so the reason why we did
• 15:59this out of necessity is we need
• 16:01this for clinical application.
• 16:02We actually needed a whole brain partition
• 16:04to covers every piece of Gray matter.
• 16:06We can't leave things on the table.
• 16:08So now this is published and this is what
• 16:10we're going to be using for us to talk.
• 16:11Now just to convince you that
• 16:13this actually is better, right.
• 16:15So you can take a dense signal, right?
• 16:18In other words, at the level of vertex
• 16:19on a cortical surface, you can parse,
• 16:22relate it prior to computing some
• 16:23metric or you can parse relate it.
• 16:25Post right?
• 16:26So if the person relation is.
• 16:29Valid and consistent across subjects,
• 16:32then this should be better than that.
• 16:34And that's in fact true, right?
• 16:35And so this is not news,
• 16:37not Glasser has shown this
• 16:38in his work and so on,
• 16:39but it's just to convince you guys that
• 16:42these parcellation are actually very,
• 16:44very useful, not just quantitatively,
• 16:46but as a feature space reduction,
• 16:49because now you are no longer
• 16:50working with 95,000 voxels,
• 16:52you're working with 700 areas
• 16:53and now you can do some clever
• 16:55feature engineering on top of that.
• 16:57So that matters, right?
• 16:59So how do we now link this to
• 17:02molecular mechanism, right,
• 17:04very broadly speaking and so.
• 17:07So, so,
• 17:07so the way we do this and the
• 17:09way we've approached this across
• 17:11all compounds ketamine,
• 17:13single sideband as well as application
• 17:17to clinical questions is what
• 17:21are the principal organizational.
• 17:26Features, if you will, of the human
• 17:28brain or the mammalian brain in general,
• 17:30and I'd like to argue one that we know about,
• 17:33is functional specialization
• 17:34across the cortical axis.
• 17:37So we know that lower order and higher
• 17:39order areas have very distinct patterns of
• 17:41feed forward and back connections, right?
• 17:44So that's an organizational principle.
• 17:45This is a classic picture from the
• 17:47Fellman and Vanessa and publication,
• 17:49which also highlights that there is a
• 17:52hierarchy and information processing from.
• 17:55Answer to association agents, right?
• 17:57We also know from work such as this
• 18:01paper from John Murray while he was
• 18:04starting here at Yale that there
• 18:06is a difference in the spontaneous
• 18:09autocorrelated activity across areas
• 18:11from non human primate data and this
• 18:15intrinsic activity scales suggesting
• 18:17a hierarchy of functional hierarchy.
• 18:19Furthermore,
• 18:20we know that during cognitive operations
• 18:23such as working memory primary.
• 18:25Visual areas such as Mt do not sustain
• 18:29signal during the mnemonic phase,
• 18:32whereas areas such as LP FC sustain
• 18:35a recurrent reverberatory activity,
• 18:38right again highlighting distinct
• 18:40functional specialization in this
• 18:42very coarse way.
• 18:43Furthermore, we know that we can leverage.
• 18:47Microstructure values from the T1 and T2
• 18:51maps to derive a proxy of myelin cortically,
• 18:56which shows a hierarchy.
• 18:58It smoothly varies from
• 19:00association to sensory areas.
• 19:02And this is also true in the macaque,
• 19:05right?
• 19:05So now we have some clues that the
• 19:07brain varies hierarchically and this
• 19:09shouldn't be controversial, right?
• 19:11Can we leverage this information to
• 19:15understand how the effects of pharmacology?
• 19:18Which we'll get to and so.
• 19:20The the motivation for this work was
• 19:23driven by a grad student in John's lab,
• 19:26Josh Burt, who did some really
• 19:28elegant gene expression mapping.
• 19:30And I'll walk you through why this matters,
• 19:31right.
• 19:31So we have this human myelin map, right?
• 19:34We've established that it has a
• 19:36hierarchical organization, right.
• 19:37And what he's just shown that when
• 19:39you use Lisas network partition,
• 19:41but in fact there is a difference
• 19:43across networks,
• 19:44right,
• 19:44that the association networks have less
• 19:47myelin than sensory somatomotor networks.
• 19:49OK, just the validity check.
• 19:51Furthermore,
• 19:51he's taken mcac myelin values,
• 19:54and he's taken tracer data,
• 19:56defining the hierarchy as feed
• 19:58forward and feedback connections in
• 20:00collaboration with David Vanessa's group.
• 20:02And he's correlated.
• 20:03This again establishing that
• 20:05there is hierarchy right.
• 20:07So far,
• 20:07so good.
• 20:08Then he went into the alien human
• 20:11Brain Atlas gene expression database.
• 20:14And he's taken from every cortical
• 20:18microarray a probe across something
• 20:20like 20,000 different genes.
• 20:23And because the amount of human
• 20:25brain Atlas has serendipitously,
• 20:27thank goodness,
• 20:28scanned every single person of
• 20:30the six people that they studied.
• 20:33Postmortem.
• 20:33We could then reconstruct the cortical
• 20:37surface right anatomy postmortem and
• 20:39map it onto the human Connectome Atlas.
• 20:43Particulate it right using the
• 20:45parcellation that I just introduced.
• 20:47Do this for every person from the
• 20:50island human brain Atlas and then
• 20:52average it to get the aggregate
• 20:54group map across every gene
• 20:57for every cortical parcel.
• 20:58Now you can imagine why this is powerful.
• 21:00Now we have an actual gene expression
• 21:02topography for every gene across
• 21:04all the areas that we have our
• 21:05new imaging maps for, right?
• 21:07So what can we do with this?
• 21:09So the first question is,
• 21:10what is the principal gradient,
• 21:12the first principal component of
• 21:15gene expression topography, right.
• 21:16And this is the picture.
• 21:18This is how it varies, right?
• 21:19And it varies this way,
• 21:22so much so that it
• 21:23did you say that this this expression
• 21:26is from 6 brains, six people. OK.
• 21:29So I'm wondering how that limits the power
• 21:31of this kind of correlation analysis.
• 21:33If you only have you have a huge number
• 21:35of genes but you only have 6 replicates
• 21:38at each for each gene in each parcel,
• 21:41that's a great question.
• 21:41So let me go back.
• 21:43So, So what Chris is really
• 21:45highlighting is something that is.
• 21:48Rental, which is you when you're running any
• 21:52analysis on the matrix of genes by parcels,
• 21:55notice that their group averaged, right?
• 21:58So what we do is we first ask, what is the?
• 22:02Coverage of that microarray probe
• 22:04in that cortical location, right.
• 22:07Is there a good signal?
• 22:08And then we evaluate the differential
• 22:10stability across individuals, right.
• 22:12So we want to be confident that it's
• 22:14consistent across these six people,
• 22:15right, and that there's good coverage.
• 22:17Then we produce a single value single number,
• 22:20which is the average right
• 22:21across these six people.
• 22:23All the analysis are done on the data
• 22:25object that's you're seeing here,
• 22:27which is at the group parcel level.
• 22:29In other words,
• 22:30the principal component does not consider.
• 22:32Variation across people,
• 22:34it considers variation across areas,
• 22:36right,
• 22:37with the assumption obviously that
• 22:39people are consistently expressing
• 22:40these genes in these areas and
• 22:42that's an empirical question,
• 22:43right?
• 22:44One that we keep talking to NIH
• 22:45and the on human brain that was
• 22:47folks that they need more brains.
• 22:49They need to produce this kind of
• 22:52mapping a sample of sufficient size
• 22:54that we can interrogate whether
• 22:56there is true human variability.
• 22:59And furthermore,
• 23:00you could imagine this matters for
• 23:02psychedelic cardiology which is if your 5 HD.
• 23:04Receptors,
• 23:04which I'll talk about in a second,
• 23:06very differentiated across people.
• 23:07One could argue that some of the
• 23:09conclusions that will draw are incorrect,
• 23:11but I think you're,
• 23:12you're you're as always always go
• 23:15to the key intuition right away,
• 23:17which is that these maps are
• 23:19average across people and therefore
• 23:22limits generalizability across
• 23:23the the entire population, right.
• 23:26We don't know
• 23:27and also implies that your principal
• 23:29components are dependent on multiple genes.
• 23:31They're looking for patterns
• 23:32of multiple genes across areas
• 23:33they're not going to give. Exactly.
• 23:35Whereas if you had a much larger data set,
• 23:37you could look for components that
• 23:39are for grades within single genes,
• 23:41but you can't do that in this data set, so.
• 23:44Exactly, exactly.
• 23:45So you could ask the question of is
• 23:48there a gradient of a single gene across
• 23:51people in an area or across areas, right.
• 23:54That's another level of variance
• 23:55that is left on the table.
• 23:57But the first question,
• 23:59the only question really that we could ask
• 24:01was what's the spatial gradient, right?
• 24:03What is the spatial topography of the way
• 24:06these genes vary across cortical areas,
• 24:08right. And this is how they vary.
• 24:11And so it turns out that that explains
• 24:14almost 30% of all the variants, right?
• 24:16Not, not everything, but a lot, right?
• 24:18And so now you could say,
• 24:20oh, what does this look like?
• 24:21What kind of looks like Milan, right?
• 24:23So you could quantify that, right?
• 24:24And it turns out it looks a lot like my own.
• 24:27In other words,
• 24:29it varies along a hierarchy, right?
• 24:31Almost 1/3 of all the variants in
• 24:34human gene expression in the adult
• 24:36brains with this these six people
• 24:39varies along the principal axis,
• 24:41where it's high in sensory
• 24:44somatomotor or low.
• 24:46And low or high in association cortices,
• 24:50right?
• 24:51This is this gradient can go both ways,
• 24:53right?
• 24:54And this is cool,
• 24:55because now you can imagine this is
• 24:57the key way that the brain breaks,
• 24:59which is across this cortical hierarchy
• 25:00and its pattern of gene expression.
• 25:04Can you then look at what genes
• 25:07are contributing substantially to
• 25:09that first principal component?
• 25:11Because you would predict that,
• 25:13for example, oligodendrocyte genes
• 25:16might contribute significantly.
• 25:17So if oligodendrocyte genes covary
• 25:19with the myelin density map, that's.
• 25:21A different way of measuring
• 25:23exactly the same thing.
• 25:25It's not telling you something new as
• 25:26opposed to if there are other neural
• 25:28genes that might be less intuitive
• 25:29that they should vary and that
• 25:30might be telling you something new.
• 25:31So can you dig into the contributors
• 25:33to this component and start to
• 25:35make that kind of inference.
• 25:36Yeah of course you can check what's
• 25:38the loading of each gene onto this
• 25:40component and so on so definitely so.
• 25:42So I don't know if Josh has that
• 25:44in one of the tables are not
• 25:45in the paper but we can check.
• 25:47It's a really good question right and and
• 25:49allows for another access to exploration
• 25:51but the point that you're already.
• 25:53You know, your your questions and and
• 25:55suggestions are already highlighting
• 25:57this which is the cortical gene
• 25:59expression variation is dominated,
• 26:00right, by a single principle axis
• 26:02which is highly correlated with these
• 26:05expressions of hierarchy and that's great.
• 26:07OK, that's an observation, right.
• 26:09And these gradients of micro scale
• 26:11properties then can contribute perhaps
• 26:13to sensory association specialization,
• 26:15but furthermore may contribute
• 26:17to the effects of pharmacology
• 26:18across these cortical areas, right.
• 26:21And so this is this is work.
• 26:23That we've published on and actually
• 26:25I'm I'm also really proud of this
• 26:27John and I Co wrote a patient with
• 26:29our colleague Bill Martin who is
• 26:30now head of Global Neuro at J&J
• 26:33and this was just recently awarded.
• 26:35So now we have a patent on this,
• 26:36not with that useful for anything
• 26:38but was it was something that we
• 26:41wanted to develop and really kind
• 26:43of drive forward anyway so.
• 26:46Sharp left turn.
• 26:47How does this relate to pharmacological
• 26:50effects of any kind and psychedelic
• 26:52psychedelics in particular?
• 26:54You guys are probably thinking when is he
• 26:55going to get to anything psychedelic related?
• 26:57Like why am I listening to this?
• 26:58So, so we're we're getting there, so.
• 27:03OK, this map and and I had some
• 27:07set up slides, but,
• 27:08but I actually want to go a little faster,
• 27:09right.
• 27:10So this is a paper published by
• 27:12Katrine in elife a couple years ago,
• 27:15right.
• 27:15And So what she's done is
• 27:17basically giving people a.
• 27:22A pill of LSD which targets the
• 27:25serotonin receptor versus placebo.
• 27:28And what we've done is we've computed
• 27:30a map of unresting state of the effect
• 27:33of LSD on every single person and this
• 27:36is the average across all the people,
• 27:39right?
• 27:41In the effects of LSD on a metric that
• 27:43we call global brain connectivity.
• 27:46So let me unpack this a little bit
• 27:48so you so you develop an intuition.
• 27:50Every warm area that you see here
• 27:52is an area that has an elevation
• 27:56in its brain wide covariation.
• 27:58When people are given LSD relative
• 28:00to cell cycle, so in other words,
• 28:03you could you could say it's
• 28:04a hyperconnectivity, right?
• 28:06But I don't like to use that term
• 28:08before impacting it first, right?
• 28:10So for instance,
• 28:11the visual cortex here would.
• 28:14This the interpretation is that
• 28:17LSD elevates the connectivity with
• 28:19the rest of the brain for visual
• 28:22cortex bilaterally and for sensory
• 28:25somatomotor cortex and MTV.
• 28:28But it reduces connectivity in
• 28:31these association areas, right?
• 28:34So this is important, right?
• 28:35In other words,
• 28:36there is a bidirectional effect
• 28:38of LSD versus placebo on sensory
• 28:41versus association regions, right?
• 28:43At least it appears that way,
• 28:45right? And so we can now isolate the Type
• 28:491 error protected effect right by doing
• 28:52TFCC protection at the whole brain level.
• 28:55And now you can see this values,
• 28:57this is what's pulling out the values
• 28:59out of this area now what Katrina has
• 29:02also done which is. Pretty clever.
• 29:05She's given people ketanserin prior to the
• 29:08administration of LSD right and contains.
• 29:11Azrin is thought to be a very selective
• 29:14antagonist of the five HT 2A receptor.
• 29:17And you'll notice that when people
• 29:19are pretreated with ketanserin,
• 29:20they look just like.
• 29:23Placebo right.
• 29:24In other words,
• 29:25there is no effect of LSD in those areas.
• 29:27And this is true for the areas that
• 29:29show a reduction in connectivity
• 29:31by LSD and the regions that show
• 29:33an increase in connectivity by LSD.
• 29:36So there is a full blockade of the effect
• 29:38of LSD by Captain Strand on average, right?
• 29:41So that's pretty cool, right.
• 29:43And and I remember showing this to to John
• 29:47Crystal years ago when we first saw this
• 29:50effect and saying like look the two maps,
• 29:53the LSD versus placebo and the
• 29:55LSD plus captain string versus
• 29:57LSD are super correlated.
• 29:58Look, they're almost the same.
• 30:00And and I'll show you what they
• 30:02look like side by side, right.
• 30:04This is same people, two different days.
• 30:07One day they're given LSD
• 30:08alone versus placebo,
• 30:10the other day they're given
• 30:11contains trend prior to.
• 30:12Was the and then contrasted to placebo.
• 30:16And I'm like, this is incredible.
• 30:17They look the same.
• 30:18And he just laughed at me.
• 30:19He was like, ohh,
• 30:20of course it's pharmacology.
• 30:21It has to be.
• 30:22Was like,
• 30:22well,
• 30:23I'm so glad it's so obvious to you
• 30:24that that you're going to get such a
• 30:27correspondence in the within subject
• 30:28effect of pharmacological effects on
• 30:30brain imaging at the surface level.
• 30:32Like I wouldn't have guessed that,
• 30:34right, but this is important,
• 30:36right,
• 30:36because it highlights that we can
• 30:39leverage surface based topography
• 30:40as an index of the effect of
• 30:43pharmacology on the human brain, right.
• 30:45And so now, now we get this map,
• 30:48this delta map, delta GBC,
• 30:50right and now we ask.
• 30:52OK, what about gene expression patterns,
• 30:54right?
• 30:54If this is truly related to the
• 30:56serotonin 5H2 receptor which we
• 30:58believe contains from this blocking?
• 31:00Then presumably the 5H2 receptor map
• 31:03ought to look like this map, right?
• 31:05So that's what we tried.
• 31:07We took the gene expression map,
• 31:08right,
• 31:09and we computed the correlation
• 31:11right across these two, right.
• 31:13And we also took some other target genes,
• 31:16right,
• 31:16that are thought in the literature
• 31:19to be targeted by LSD.
• 31:22And then we computed a similarity of
• 31:24correlation between this and this is
• 31:26serotonin is right up here, right.
• 31:28So it's the of these, it's the most.
• 31:31Positively correlated.
• 31:32And then we repeated this for the
• 31:34entire distribution of all genes
• 31:36from the alien human brain Atlas.
• 31:38And again,
• 31:39serotonin comes out here in the
• 31:4195th at almost 96 percentile, right?
• 31:44So there are some things that
• 31:45by chance could
• 31:45be higher, but this is pretty.
• 31:47Encouraging as a initial proof of
• 31:49principle that we can actually map
• 31:51human gene expression in relation
• 31:52to the pharmacological effects on
• 31:54the human brain in vivo of a given
• 31:57pharmacological agent that was blocked by
• 32:00the hypothesized receptor antagonist, right?
• 32:03That's in my mind pretty cool.
• 32:06And so, so you also see the opposite,
• 32:09which is HR7, which in my mind has
• 32:11some very interesting pharmacology,
• 32:13but we won't get into that today, right?
• 32:15It's just you also see the opposite
• 32:17effects in this, this kind of analysis.
• 32:19OK. So. How do we take
• 32:23this quick question?
• 32:24If you go back, I'm thinking about
• 32:27the 1A receptor which is usually
• 32:29pre synaptic on Axon terminals,
• 32:31which means that the receptor is
• 32:34not in the same place as the M RNA.
• 32:37That's what you're doing here.
• 32:39When you're when you're looking at the
• 32:41distribution of gene expression across
• 32:43the brain, you're looking at M RNA,
• 32:45which is basically where the cell bodies.
• 32:49Right and. That should be pretty good
• 32:52because it's amended and Reddick.
• 32:53But for the 1A receptor there,
• 32:55there's going to be a substantial
• 32:57dissociation between where the M
• 32:58RNA is and where the receptor is,
• 33:00which I don't think you can get
• 33:01it at this technique.
• 33:02So just for some receptors that may be
• 33:04applicant, you're absolutely right.
• 33:06So you're absolutely right.
• 33:08So what you're highlighting is the nuance,
• 33:11the Super important nuance between
• 33:13the ligand and the receptor, right.
• 33:15And basically the fact that the M RNA
• 33:16may be coding for the ligand or the
• 33:18receptor and in the cases where it's
• 33:20coding for the receptor, this analysis.
• 33:22Approach may be very informative,
• 33:24but in the cases where it's
• 33:25coding for the ligand,
• 33:27it may or may not be informative.
• 33:28It's not about whether it's the ligand,
• 33:29it's where the receptor is on the neuron.
• 33:33So for pre, for pre synaptic receptors,
• 33:36for receptors that are
• 33:37targeted to Axon terminals,
• 33:39we see this in the animal literature
• 33:41all the time where the receptor
• 33:42in the M RNA are in different
• 33:43places because the receptor is
• 33:44way out on the Axon terminals,
• 33:46which is in a different part of
• 33:48the brain from the cell body.
• 33:49It's not about the ligand,
• 33:50it's about where the receptor
• 33:52is localized in the neuron.
• 33:53So you're highlighting something
• 33:55even different, right, which is which
• 33:56is now beginning to appreciate it.
• 33:58So you're saying that the postsynaptic
• 34:01receptor expression of the five HD.
• 34:03Way. Is potentially captured very well
• 34:07by the M RNA and the the the probes
• 34:12whereas the presynaptic 1A may may be a
• 34:15very different phenomenon because it's
• 34:17on the presynaptic terminals right.
• 34:19And therefore you're not binding to it.
• 34:21And then furthermore furthermore it's
• 34:23this is where my thinking was going
• 34:26which is the ligand versus the receptor,
• 34:29right because now you have that third
• 34:31axis of variation and so and and to your
• 34:33point this gets complicated because.
• 34:35When you give a substance like a psychedelic,
• 34:38you have post,
• 34:39you have polysynaptic distal effects, right?
• 34:43Which is it's going to travel right through
• 34:46the Axon and potentially shuttle onto the.
• 34:50Receptors and activate those terminals
• 34:52on distal neurons that are not in the
• 34:55local tissue bed of the high expression,
• 34:57which is why and this is actually
• 34:59gets really nuanced why looking at
• 35:01the dense GBC at the voxel level and
• 35:04partially the GBC doesn't necessarily
• 35:06fully map onto one another because that
• 35:09level of granularity begins to matter,
• 35:11right.
• 35:11And you can now appreciate right
• 35:13that if you are averaging signal
• 35:15within an area or if you're looking
• 35:17at boxing level pharmacology,
• 35:18right and furthermore if you can.
• 35:20Some even deeper into
• 35:22columnar level pharmacology,
• 35:23but they're an important ones here, right?
• 35:25But at the very coarse level,
• 35:28you can at least begin to
• 35:30identify these first principles,
• 35:31which is pharmacological
• 35:33neuroimaging topographies with GBC,
• 35:35which is this random freaking measure, right?
• 35:37That that seemingly has these
• 35:39properties that we really like maps
• 35:42onto gene expression gradients, right?
• 35:44Like who would have thunk it?
• 35:46And it's not.
• 35:46And and the important thing about this
• 35:48is that if you use some graph, theoretical.
• 35:51Distraction.
• 35:51Without a surface map,
• 35:53there's no freaking way you can
• 35:55get this right.
• 35:56And that's actually kind of the take
• 35:58away that I was trying to get at right
• 36:00is that that this is obscured if you
• 36:02do not have a surface map, right?
• 36:04You just can never see it, Umm,
• 36:07because that's what drives the
• 36:10correspondence, right, and the location.
• 36:12So anyway, so, so,
• 36:13so now I I want to be sensitive to time,
• 36:16so I may have to kind of speed
• 36:17up to some of this stuff.
• 36:18So, so basically the point of this is
• 36:20how do we map this onto neurobehavioral,
• 36:23geometric models in in clinical
• 36:26population level analysis,
• 36:27right.
• 36:28And so this cartoon is just
• 36:29highlighting that there's some
• 36:31brain to behavioral relationship and
• 36:32that it's probably or some oblique,
• 36:34maybe even not linear.
• 36:35And we don't know what it is.
• 36:37And so this is work really purely done
• 36:41by my former student now research
• 36:43scientist here in our department,
• 36:45Lisa and and she's published
• 36:48this in life earlier last year.
• 36:51After a whole saga.
• 36:53So,
• 36:54so I I I want to this gets a little
• 36:56technical, so I'll try to kind of keep
• 36:58it clear and then get to the key points.
• 37:01So if we're going to leverage
• 37:03pharmacological new imaging that's
• 37:05even benchmark with gene expression or
• 37:07what have you right to achieve brain
• 37:10behavioral models that can actually be
• 37:12deployed for any therapeutic purpose.
• 37:14There are some criteria that I'd like to
• 37:16argue we need to really hold in mind, right.
• 37:19And these criteria are are not exhaustive and
• 37:21things that I've learned the hard way that,
• 37:24you know, if you don't do this,
• 37:26things are just brittle.
• 37:27So the first one is more
• 37:29kind of for the whole field,
• 37:30which is that anything that we
• 37:32produce as a field has to scale and
• 37:35interoperate in an informatics way.
• 37:38So for instance,
• 37:39what Pronet is doing and what Professor
• 37:41Woods is doing in our department with
• 37:43this massive worldwide consortium.
• 37:45Right.
• 37:45So, so the days of me working on my PC
• 37:49and producing imaging are gone, right?
• 37:51Like it's no longer that.
• 37:52So then the measure selection,
• 37:55and I mean the behavioral measure selection
• 37:58matters here and it matters a lot.
• 37:59And I'll show you why. Then.
• 38:02How do we cross validate those
• 38:04behavioral models,
• 38:05right.
• 38:06And this is stuff that has to do with
• 38:08trustworthiness of the reproducibility
• 38:10of those models at the behavioral level.
• 38:14No imaging yet.
• 38:15This is this is something that that
• 38:17we also found out is very important.
• 38:20Then Criterion 4 is how do we then
• 38:22produce a robust and interpretable
• 38:25neuroimaging maps that are linkable
• 38:28to that behavioral variation?
• 38:29And then finally,
• 38:30how do we cross validate that
• 38:32brain behavioral?
• 38:33Right.
• 38:33So this is a lot of stuff to cover
• 38:35in like 15 minutes.
• 38:36So some of the hit on some of the highlights
• 38:39and then we'll pause your questions.
• 38:41So this is.
• 38:42You know,
• 38:43a trivial point, right,
• 38:44it's just hard to do,
• 38:46which is we need informatics
• 38:47solutions that scale.
• 38:48And Yale is at the forefront of this.
• 38:50I think that what we're doing
• 38:52in our department,
• 38:53I'm tremendously proud of and I
• 38:54think some of the work of faculty
• 38:56on this call and others like we're
• 38:58really pushing the boundary of this.
• 39:00This is the point is really things
• 39:01have to scale and drop rate if
• 39:03we're going to develop precision
• 39:05medicine solutions, right.
• 39:06So I'll just forward and just
• 39:08say a key thing inside this
• 39:10architecture is analytic discovering.
• 39:12Years.
• 39:13And so this is the workflow from
• 39:15Lisa's paper,
• 39:15just as a shameless plug.
• 39:17But the point is that analytics have to be
• 39:20organically built into this for the Discovery
• 39:22science engine to work where it cannot be.
• 39:25Data collection devoid of analytics.
• 39:27It's it's it's all,
• 39:29it's all combined, right?
• 39:30So talk through how we do this with
• 39:32a particular analytic framework using
• 39:34a data set that's called Beast Snip.
• 39:37Our colleague,
• 39:38Godfrey Pearlson was one of the
• 39:41principal investigators on the original.
• 39:43Snip that made it into the public domain,
• 39:44and now they're on to be snipped too.
• 39:46I don't even know three,
• 39:47but this is a public domain
• 39:49datasets that made it into NH.
• 39:51We downloaded it out of the National
• 39:53Debt archive and processed it using the
• 39:56human connectome processing pipelines.
• 39:58Like that's as much as I'll
• 40:00say to speed ahead, OK?
• 40:01So I've shown this several times.
• 40:04Maybe some of you have seen these data,
• 40:06but now it's published and I can kind
• 40:07of show you the full gamut of this.
• 40:09So the first question that we asked
• 40:13ourselves was what is the covariance
• 40:16structure across people in the symptom
• 40:20geometry of the psychosis spectrum
• 40:22population of these 436 people?
• 40:25And you're looking at pans,
• 40:27these are pans items, right?
• 40:28The backs is on top backs,
• 40:31which is the brief.
• 40:32Assessment of cognition and
• 40:34then pans items here.
• 40:35So this is the covariance matrix
• 40:38and there's some correlation between
• 40:40these right across 436 people.
• 40:42So in other words,
• 40:43they're structure between these symptoms,
• 40:45which is expected.
• 40:46This is not new, this is, this makes sense.
• 40:49But when you plotted across
• 40:51the DSM categories, right,
• 40:52where bipolar is shown in yellow,
• 40:54schizoaffective in this
• 40:56kind of orangish color,
• 40:58dark red is schizophrenia and all pro bands,
• 41:00all patients are shown in black,
• 41:02you know, I like to argue that.
• 41:06There's a lot of variation in each one of
• 41:09these sub scores on the pans and cognition,
• 41:12but not really clear,
• 41:13very clear distinctions between
• 41:15diagnostic categories, right?
• 41:16And so this is not news.
• 41:19Psychosis spectrum disorder is heterogeneous,
• 41:20exhibits covariation symptoms
• 41:22across clinical scales. OK, great.
• 41:24You know, cool story Allen.
• 41:26So now what?
• 41:26So the question is what is the
• 41:28dimensionality of this solution,
• 41:30right.
• 41:30Is there a low dimensional solution that
• 41:32we can reduce the map to the brain right.
• 41:34Can we do that and simply right.
• 41:37You could ask is there say principal
• 41:40component analytic solution that
• 41:42explains these data or a factor analytic
• 41:44solution or K means clustering solution,
• 41:47just something that is looking at
• 41:49the covariance structure of the
• 41:51data in a lower dimensional space.
• 41:53And so it turns out yes.
• 41:55Right.
• 41:55So these are the components that come out,
• 41:57which brings me to criterion too, right?
• 41:59Can we select the right measures
• 42:01to map onto the brain? Right.
• 42:03And can we obtain an interpretable
• 42:05solution here?
• 42:06Right.
• 42:06So, I mean, I'm not going to walk
• 42:08through these principal components.
• 42:09More importantly,
• 42:10when I want to highlight is what the geometry
• 42:12looks like just so you can get an intuition.
• 42:15Every dot is a patient.
• 42:16They're color-coded as noted here,
• 42:18right?
• 42:19These arrows in this space are
• 42:22vectors of the pans and backs.
• 42:28Average scores. The green arrow is backs,
• 42:32which is almost perfectly
• 42:33collinear with the cognition axis.
• 42:35That's one of the principal components,
• 42:37the black dots that you may see
• 42:40here healthy controls right?
• 42:41And then the two the the the
• 42:45arrows that are coming up the
• 42:48the blue and the purple are the.
• 42:52Negative and positive symptoms,
• 42:54respectively, right?
• 42:55Notice that they project onto
• 42:58this access under an angle.
• 43:01It's not. They're not collinear,
• 43:03and they're obliquely rotated, right?
• 43:06And then there's a global dysfunction
• 43:09which is all the patients are not
• 43:11functioning as well as controls.
• 43:13So it's this PC three that we're
• 43:15going to talk about, right.
• 43:17So this is a static picture of that solution,
• 43:20right.
• 43:21And if if you can see after correct
• 43:24schizoaffective is misspelled,
• 43:26but if you can see the positive
• 43:28and the negative vectors,
• 43:29they form a 45 degree angle onto PC-3.
• 43:32OK.
• 43:33So now what is this action you
• 43:35look like numerically, right?
• 43:36Zoom in and these are the linear combinations
• 43:39and this solution right in this sample.
• 43:42So if I plot the DSM categories,
• 43:44you'd say there's an effect.
• 43:47Right. They don't differ,
• 43:49but that's actually the point.
• 43:51The point is that when you cut
• 43:54through this geometry with the.
• 43:56Data-driven solution.
• 43:57You ought not to see differences in DSM
• 43:59categories because they don't seem to
• 44:02actually follow natural variation, right?
• 44:03And so how can I convince you of that, right?
• 44:06So.
• 44:06So let's take a look at the
• 44:09configuration of the PC-3 items, right.
• 44:12So typical person would be somewhat
• 44:14delusional, conceptual, disorganized.
• 44:16They are hallucinating,
• 44:17they have some excitement,
• 44:19grandiosity, right?
• 44:20But they're not, you know,
• 44:22purely collinear with the negative symptoms.
• 44:24They have something, some they don't.
• 44:26They're a little bit cognitively impaired,
• 44:27right?
• 44:28But again not a clean, you know,
• 44:31one to one mapping between these these axes,
• 44:33right, between the the subscores of the pans.
• 44:36So again, you know,
• 44:37let's put this to the test
• 44:38I I'm a competitive person,
• 44:39I like competition, right.
• 44:40And and I like to, you know,
• 44:43create competitions,
• 44:44incentive questions, right.
• 44:45So let's see is the SM going to
• 44:48outperform a data-driven solution, right.
• 44:50Because we need it to map it
• 44:52onto pharmacology, right?
• 44:53Like we need something that is robust, so.
• 44:58Criterion 3.
• 44:59Before we even get there right,
• 45:01we have to check that the
• 45:03solution of this model is stable.
• 45:05So this is a summary of the leave
• 45:07each site out cross validation,
• 45:10a summary of the predicted
• 45:11versus observed scores,
• 45:12a summary of the predicted versus
• 45:15observed single subject scores
• 45:16from K fold bootstrapping,
• 45:18and similarity of the actual loadings
• 45:20on the PCA for leave site out five
• 45:23fold bootstrapping and split half.
• 45:25And hopefully this shows you that
• 45:27the solution is really stable.
• 45:28Which means that the basement consortium
• 45:30did a really good job actually,
• 45:32right.
• 45:32They collected and and asset clinically
• 45:34the data in a very consistent way
• 45:37and we're able to get a pretty
• 45:39good stable behavioral model,
• 45:41right. So the PCA variance is generalizes,
• 45:45the score is generalized and the
• 45:46PC weights generalized, right.
• 45:48Otherwise why are we mapping it onto
• 45:50the brain if it doesn't, right? Cool.
• 45:52So now let's actually go even further
• 45:54from DSM and take pans positive symptoms.
• 45:57Let's give pans a fair shot because
• 45:59this is what the industry uses, right?
• 46:01If you're going to test if
• 46:03a antipsychotic works,
• 46:04you're going to use pans positive symptoms.
• 46:06That's your benchmark.
• 46:07That's the gold standard for the industry,
• 46:09right. And and this is the
• 46:13psychosis configuration PCA effect,
• 46:15which required only only one level
• 46:18of supervision, which is to pick PCA.
• 46:21That's it. We just said.
• 46:22Let's run a PCA.
• 46:23So now, which one will give you
• 46:25a better brain map, right?
• 46:26That's what we care about.
• 46:28We care which Brain Mac is better.
• 46:31And so again,
• 46:31we're going to use GBC as the
• 46:33brain measure and we're going to
• 46:36calculate the variation from each.
• 46:37Parcel to every other parcel using this
• 46:40quantitative technique that I explained,
• 46:41right.
• 46:42And so the intuition again is that we're
• 46:44going to get this value for every parcel.
• 46:47We're going to then correlate
• 46:49the area level signal with the
• 46:52symptom for every patient.
• 46:54And then we're going to get in a
• 46:56cross subject map that tells us
• 46:58how people vary across the sample
• 47:00with respect to their GBC, right.
• 47:02So it's an individual difference analysis.
• 47:04And so this is the map you get with pans
• 47:07with 436 people and this is the map.
• 47:09You get when you use the PC-3.
• 47:12Now.
• 47:12I want to just pause here because
• 47:15hopefully it's self-evident to
• 47:17everybody which one is better
• 47:19and if somebody says a.
• 47:21I I hope they're joking.
• 47:24So.
• 47:24This is not nothing done except simply
• 47:27taking a data-driven behavioral
• 47:30analysis of pens and backs,
• 47:33a data-driven neural measure with
• 47:35the only piece of supervision being.
• 47:38Reduction of the FC matrix using GBC.
• 47:41And then you get this slice through
• 47:43the geometry that hopefully one
• 47:45could argue is is better and
• 47:47quantitatively it is better, right?
• 47:49You can actually check that statistically
• 47:51that the variance covered is higher
• 47:53and that the range of the Z values is better.
• 47:56You can do all sorts of other things,
• 47:57but it's just better.
• 47:59So, OK, now we have something right
• 48:03now criterion 5 is,
• 48:05is this thing generalizable?
• 48:07And what I mean by that is if I were to.
• 48:11Say Mark or Chris,
• 48:13can you guys use the weights,
• 48:15the actual thing that I found
• 48:18here and reproduce the exact map?
• 48:21Using a split half cross validation
• 48:24of the model,
• 48:25can you get the same picture again?
• 48:28That's that's what we care about.
• 48:29Not just that you can point to five,
• 48:31reject the null and publish,
• 48:33but that the picture of the neural
• 48:36topography is reproducible.
• 48:37And this is what we get in this case,
• 48:39right?
• 48:40So and we did this 10,000 times,
• 48:43but half and various ways and tried to
• 48:46break it. You know, pretty robust.
• 48:48Both dense level and the parcel level, right?
• 48:52And this is only 219 people, right?
• 48:54We're not talking gargantuan samples here,
• 48:56right? It's just that you have the
• 48:59right slice through the geometry and
• 49:01then all of a sudden you're getting
• 49:03maps that reproduce even in patients.
• 49:05So now this is the engine that we
• 49:08submit this to in order to select
• 49:11the most stable features, right.
• 49:13And I'm not going to walk through
• 49:15this because it's certainly dense,
• 49:16but it's, it's simply, you know,
• 49:19under the hood it's some basic
• 49:21linear algebra of optimizing each
• 49:23feature in relation to stability
• 49:25criteria from the out of sample.
• 49:28Generalization.
• 49:28And so then you can ask the question of which
• 49:32parcels of the map should we trust the most?
• 49:35And those are the parcels that
• 49:36then we can use as a, you know,
• 49:38for further feature engineering.
• 49:40Interestingly though,
• 49:41what you can also do is then once
• 49:43you've done this and you find
• 49:45your trustworthy parcels right,
• 49:46the ones that truly generalize,
• 49:48you can then ask how do they covary in
• 49:51relation to the behavioral feature selection?
• 49:54Turns out there's a nonlinear relationship,
• 49:56right?
• 49:56Which means that the more extreme the
• 49:58person is on their behavioral loading,
• 50:01the more you trust their neural net.
• 50:03That makes sense, right?
• 50:05That's intuitive.
• 50:06And so when you then do this and purely
• 50:09filter people based on behavior,
• 50:11just take the 10th.
• 50:13And the 90th percentiles of the
• 50:16behavioral scores and segment that way.
• 50:19Then you can begin to segment
• 50:22based on neurobehavioral similarity
• 50:25of the map until you get.
• 50:27To the very peak of this patient
• 50:29selection and then you ask how accurate
• 50:31is this model and then you can see that
• 50:34it's pretty damn accurate at a sample,
• 50:36right.
• 50:36So in other words it's classifying people as
• 50:38plus or minus and in terms of their range.
• 50:41And then you can repeat this on
• 50:43a completely independent sample
• 50:44and again show that it works.
• 50:46In terms of segmentation by the way,
• 50:48Chris,
• 50:49this is the OCD and and skids data
• 50:53set which is cross diagnostic that
• 50:55we tried this with, right so.
• 50:57This isn't even patients schizophrenia
• 50:59anymore right.
• 50:59It's just saying does your brain
• 51:01map look like some norm that we
• 51:04can behaviorally incur right.
• 51:06So it's really about symptom
• 51:08configurations right.
• 51:09No longer about our you know,
• 51:11do you have a diagnostic category
• 51:13in it anyway,
• 51:14how can we leverage gene expression
• 51:16out to molecularly benchmark this
• 51:18and link it back to pharmacology.
• 51:20So again I'm going to just remind you
• 51:21of this framework Gemini dot, right.
• 51:23So now we can take this PC three
• 51:26map that is trustworthy.
• 51:28And again,
• 51:29correlated with gene expression patterns
• 51:31in the same way that we've done with MSD.
• 51:33And this is just proof of principle,
• 51:35right?
• 51:35Again, we can show some relationships,
• 51:37I'm not going to get into this too much,
• 51:39but for instance,
• 51:40you can show that the interneuron markers
• 51:43or GABA subunits or serotonin
• 51:45receptor subunits have
• 51:47correspondence with this map.
• 51:48It's I'm not claiming mechanism or anything,
• 51:52I'm just saying you could do this,
• 51:53right. This is doable.
• 51:55But finally to conclude, you can then.
• 51:59Benchmark this against our
• 52:01pharmacological targets.
• 52:02So this is actually an in vivo ketamine map,
• 52:06same GBC measure,
• 52:07healthies versus healthy people,
• 52:09placebo versus.
• 52:13Infusion, right?
• 52:14And then we can select people along the
• 52:17axis that presumably varies in relation
• 52:20to that work without ever optimizing it.
• 52:23We're not optimizing it yet.
• 52:24And then you take two people
• 52:25on the extreme ends, right,
• 52:27and these are their actual brain maps.
• 52:29These are two people diagnosed
• 52:30with schizophrenia, right?
• 52:31They both have the same diagnosis.
• 52:32Yet I'd like to argue that their symptom
• 52:35configurations are completely different
• 52:36and their brains don't look the same.
• 52:38Right. Yet we're treating them the same.
• 52:40We're giving D2 blockers to both of these
• 52:42people as the initial line of defense,
• 52:44when in fact, who knows,
• 52:46maybe one person would respond
• 52:47way better to close the people.
• 52:48And we have no idea that that
• 52:50is true or not true, right?
• 52:52But you can then quantify that using.
• 52:56This.
• 52:56Framework that Lisa has advanced in
• 52:59relation to a given target and you can
• 53:01say which person is more similar, right?
• 53:04So this person looks like PC-3,
• 53:06which looks like ketamine.
• 53:08So presumably this person will get
• 53:09worse if you give them ketamine and this
• 53:11person would maybe even get better, right?
• 53:13I don't know.
• 53:14But you could do the same thing with
• 53:17LSD now and repeat this for another
• 53:19access and recapitulate this principle.
• 53:22Now these maps can be iteratively optimized.
• 53:27This is a.
• 53:28Feature selection problem.
• 53:29Now we can use the LSD,
• 53:31psychedelic or ketamine target
• 53:33maps to find people who may
• 53:35benefit the most quantitatively,
• 53:37rationally,
• 53:38iteratively in a fast fail algorithm that
• 53:41says these two patient populations ought
• 53:44to show the opposite effects of this drug.
• 53:47And that's a strong inference
• 53:48rational framework, right?
• 53:49And so just to summarize,
• 53:52like I do think we need
• 53:54informatics and scalability.
• 53:55I do think we need to first and foremost.
• 53:58Their behavior,
• 53:58right?
• 53:59Select the right combination we need to
• 54:02achieve criteria for trustworthiness
• 54:04of those behavioral models, right?
• 54:07That's that's a must.
• 54:09Then and only then do we go to brain imaging,
• 54:12and then we need an interpretable.
• 54:14Robust and generalizable solution of
• 54:17the brain back which in turn then we
• 54:20can cross validate with pharmacological
• 54:23and gene expression and other metrics
• 54:26that that the field is bringing to bear.
• 54:28So, so in summary,
• 54:29I I do think that we have an
• 54:31opportunity here, right.
• 54:32I think that what we're doing in our
• 54:34department is truly transformative.
• 54:35I'm just one out of many people who
• 54:37are doing this work and we have I
• 54:40think an iterative framework right for
• 54:42really dissecting heterogeneity with.
• 54:44Imaging and behavior.
• 54:45And this can be optimized actually
• 54:47again for patient selection and
• 54:49precise delivery of of psychedelic
• 54:51compounds to the right patient.
• 54:53So I'll stop there and think
• 54:55I'm actually in time.
• 54:56Remarkable.
• 54:58Impressive. Thank you, Alan.
• 55:02That was a remarkably lucid presentation
• 55:04of some very complicated material.
• 55:09Questions. Comments. We have just a
• 55:12couple minutes before official end time.
• 55:25Ellen, I wonder if you could
• 55:27swing back to speculate.
• 55:29Since sort of the motivating you,
• 55:32you've spent a lot of time talking
• 55:33about the framework, the technology,
• 55:34the analytics and the potential and and
• 55:37just a couple slides on the psychedelics.
• 55:39In the middle,
• 55:40which I think is fine because, you know,
• 55:43it's important for us to recognize this.
• 55:46What what what you're working on,
• 55:47but I wonder if you could project OK.
• 55:50This group is motivated primarily by an
• 55:52interest in how do the psychedelics work?
• 55:54Who can they help? You know,
• 55:56how would you imagine over the coming years?
• 55:58And I know you've thought about
• 56:00this a lot because you're doing
• 56:01it and planning on doing it.
• 56:03So how, how would you envision
• 56:05applying this framework to a deep,
• 56:07to developing a deeper
• 56:09understanding of how psychedelics?
• 56:11Affect the brain both in terms
• 56:13of their acute, you know,
• 56:15psychotomimetic, dissociative,
• 56:16whatever effects and in terms of
• 56:19their longer term therapeutic effects.
• 56:21Yeah, so.
• 56:24So there
• 56:25there's two pieces of work
• 56:26that I didn't have the time to
• 56:27highlight and I was wrestling with.
• 56:29Do I want to go into them or not?
• 56:30And so, so one paper that Katrin
• 56:33published is looking at time
• 56:35dependent effects on the brain of
• 56:37silybin and the same imaging session.
• 56:39So one thing that she's done that I
• 56:42think is really impressive as shown
• 56:44the evolving neural neural effect
• 56:46of these compounds in the same
• 56:49person and showing how these maps,
• 56:51these topographies evolve as we.
• 56:54Selecting data overtime and that
• 56:56gives us confidence of the, the, the.
• 57:01Basically neural targeting engagement.
• 57:02So that's one thing that I think
• 57:04we really need more of these neural
• 57:07targeting management and so then then
• 57:08what Josh has actually done cleverly
• 57:10in in a in a sister paper to Lisa's paper,
• 57:12which is a whole nother beast
• 57:14that I didn't want to get into.
• 57:15John senior author on that,
• 57:17he's actually taken the the observation
• 57:19from Katrina was the observation
• 57:21and then fit gene expression to the
• 57:24computational models out of John's labs.
• 57:26And I,
• 57:27I I really didn't want to get into that
• 57:29because the the technical detail behind it.
• 57:31This is maybe beyond our time scope today,
• 57:34but you guys should invite
• 57:35him to talk about that.
• 57:37And So what he's done is put in
• 57:40gradients of 5H2A pharmacology
• 57:42into the biophysical models,
• 57:44right,
• 57:45simulated surrogate models and then
• 57:47fit them to individual people given
• 57:49LSD and found that actually explains
• 57:52the data way better in relation to
• 57:54their symptoms that they get acutely.
• 57:56So that's that's another paper.
• 57:58So these two pieces of work are all
• 58:00about neural target engagement.
• 58:01Confidence.
• 58:01And then Chris,
• 58:02your question is how do we apply this?
• 58:04What do we do with this in relation
• 58:07to helping people who may benefit
• 58:09from the the administration of
• 58:11these psychedelics and that has
• 58:13to do with finding individuals in
• 58:15the general population whose?
• 58:18Purported or potential neural
• 58:21system disturbance alteration,
• 58:24whatever term you want to use
• 58:25in relation to their behavior.
• 58:27In this case mood maps onto that
• 58:31neural target engagement profile right.
• 58:34And so the question becomes there
• 58:36are two questions right that we're
• 58:39after is the effect of LSD and or
• 58:42silicide been uniform across people.
• 58:44In other words if you give it
• 58:46to me and you and mark and.
• 58:48Anita across.
• 58:49Doses.
• 58:50Is our brain topography gonna look the same?
• 58:54Turns out not,
• 58:56that's not true and that matters.
• 58:59So for patient precision
• 59:01delivery that matters, right,
• 59:02because if say you are particularly
• 59:05amenable to respond to that compound,
• 59:07but I'm not right,
• 59:08then you wouldn't give it to me.
• 59:10And that has nothing to do with
• 59:12my behavioral alteration per se,
• 59:14but it may have a lot to do with the
• 59:17receptor occupancy and the the nature of the,
• 59:20you know, individual.
• 59:21Creation.
• 59:21Turns out this is unpublished work,
• 59:23but ketamine is even more high than.
• 59:25Functional,
• 59:26right?
• 59:26Turns out that there is no one axis
• 59:28of the average effect
• 59:29Academy on the human brain.
• 59:31It's actually highly dimensional,
• 59:33which obscures paradoxically
• 59:35the average effect, right,
• 59:37if you have multiple dimensions.
• 59:38And so this is what we're after.
• 59:39We're after mapping variation of
• 59:42psychopharmacology within and
• 59:43across people in order to then
• 59:45informed precision of how it
• 59:47relates to circuit disturbance.
• 59:53I don't thank you very much for your call.
• 59:54I have a question about the study
• 59:57that you show about the functional
• 01:00:01connectivity after LSD and was it in the
• 01:00:05acute phase or like post acute or like,
• 01:00:08I just want to know how long after
• 01:00:10those of like any psychedelic,
• 01:00:12we have a question. And two scans,
• 01:00:16one at 75 minutes, one at 300 minutes.
• 01:00:18So you could argue because the ketanserin
• 01:00:21and the LSD half lives have slightly
• 01:00:24overlapping and distinct curves.
• 01:00:26So we wanted one early and one late,
• 01:00:28and it turns out that that matters.
• 01:00:31So like do you know of any?
• 01:00:33Like is the Hyperconnectivity continues
• 01:00:35after for example after one week?
• 01:00:39I don't know. We don't know.
• 01:00:41That's a wide open question, right.
• 01:00:42So. So I don't know that anybody's
• 01:00:44looked at these sustained effects.
• 01:00:46What we have some stuff from Arena
• 01:00:48Australia's data set, right.
• 01:00:49Which again I it's it's really her
• 01:00:52story to report but but it turns
• 01:00:53out also when we give ketamine
• 01:00:55and look at people a day later,
• 01:00:57right, with F MRI and behavior,
• 01:00:59there's this and you guys know this, right.
• 01:01:01There's this crazy inverted V relationship.
• 01:01:03Some people have a sustained effect of
• 01:01:06the antidepressant phenomenon and other
• 01:01:08people go right back to where they were.
• 01:01:10And we don't know why this is.
• 01:01:12This is unexplored neurobehavioral effects.
• 01:01:14We don't know,
• 01:01:15but we know that everybody
• 01:01:17acutely shows some kind of.
• 01:01:19Clinical efficacy,
• 01:01:20but then a day later you have this rebound
• 01:01:23and who is rebounding and who is not why?
• 01:01:26You know there's ideas about synaptic
• 01:01:28plasticity and LTP like phenomena
• 01:01:30and and you know which people have
• 01:01:32that dendritic proliferation would
• 01:01:34then stabilizes and who is most
• 01:01:36likely to benefit from that kind
• 01:01:39of and you know psychedelics and
• 01:01:40ketamine very different from ecology,
• 01:01:42very,
• 01:01:42very different but maybe converging on
• 01:01:44some endpoint of exciting and driving
• 01:01:46the circuits into an LTP like phenomena,
• 01:01:48we don't know.
• 01:01:49And
• 01:01:50is there any relationship,
• 01:01:52any association between the
• 01:01:53degree of this hyperconnectivity
• 01:01:56and response to treatment?
• 01:01:58We don't know that's nobody has
• 01:02:00that data set again nobody's
• 01:02:03done MD either you know major,
• 01:02:06major depression or you know severe mood
• 01:02:08disturbance data set or experiment in
• 01:02:10which people were given either ketamine,
• 01:02:12indoor silybin or randomized to one of
• 01:02:16these arms scanned prior at baseline.
• 01:02:19Scanned acutely scanned post
• 01:02:21and then scanned later when they
• 01:02:23either sustain their recovered and
• 01:02:25understood what predicts it right.
• 01:02:26Like wide open question.
• 01:02:27And I think our department is unique
• 01:02:29position to go after this right.
• 01:02:31I think there's plenty of
• 01:02:33people who have the the means,
• 01:02:35expertise and talent to go after
• 01:02:36this and it's a fascinating question
• 01:02:38like this is what we need to know.
• 01:02:41Thank you. No, my pleasure.
• 01:02:45Anahita stole my question.
• 01:02:46I I also had the question about can
• 01:02:48we predict who was going to respond
• 01:02:50if they have a certain a certain
• 01:02:52pattern that that that that shows up.
• 01:02:55But I also wonder whether there have
• 01:02:57been other neuroimaging studies,
• 01:02:58not with psychedelics but with
• 01:03:00other treatments that showed
• 01:03:02changes in the connectivity such
• 01:03:05as T or presence. Totally, totally.
• 01:03:09So actually I forgot.
• 01:03:10So charity of Donna did a nice study
• 01:03:12where he looked at meta analysis of of.
• 01:03:15Uh, effects of ketamine. I don't wanna,
• 01:03:17I actually don't wanna forget that I
• 01:03:19think John's on the paper, John Christow.
• 01:03:21So there is some evidence of this in
• 01:03:23the literature that people have done,
• 01:03:24but just not the experiments that
• 01:03:26you guys were asking about that.
• 01:03:28But to your point, yes.
• 01:03:29In fact, we worked with Anil Malhotra and
• 01:03:32to look at the effect of clozapine, right.
• 01:03:35We're actually writing this up
• 01:03:36for publication as we speak.
• 01:03:37And so, yes,
• 01:03:38in fact you can predict and and
• 01:03:39turns out that these are very strong
• 01:03:42effects actually when you get
• 01:03:43people who are responding, right.
• 01:03:45The neural maps of predicting who responds
• 01:03:48are actually quite nice and clean.
• 01:03:50It's just that these are small samples
• 01:03:53like we're talking 141520 people, right.
• 01:03:54So it's just the first wave of
• 01:03:56work that's coming out, right?
• 01:03:58Like this is the next generation,
• 01:03:59right.
• 01:04:00They don't like precision pharmacology
• 01:04:02to dissect individual variation in
• 01:04:04relation to neurobehavioral effects.
• 01:04:06Like, I I'm just super excited,
• 01:04:08right,
• 01:04:08because I think this is actually happening.
• 01:04:10Like we actually see this now like
• 01:04:11the next 5 to 10 years as possible.
• 01:04:16Great. Thanks.
• 01:04:20Very excited. I'm sorry, go ahead.
• 01:04:24I was just going to say we're
• 01:04:25overtime and I wonder if
• 01:04:27we should wrap up that if.
• 01:04:28Well, I was just going to say,
• 01:04:30I'll make it quick then.
• 01:04:31So very exciting
• 01:04:32work, great presentation and the
• 01:04:37several questions that were just
• 01:04:40asked made me think about some of our
• 01:04:43work in the psychotherapy Development
• 01:04:45Center and the data that we've been
• 01:04:47collecting over the past 15 years
• 01:04:49by integrating F MRI measures into
• 01:04:51randomized clinical trials and the
• 01:04:53potential for these approaches we've
• 01:04:55been using things like connectome based.
• 01:04:58Predictive modeling, but there
• 01:04:59are many different approaches to understand
• 01:05:03better how people may respond to treatment.
• 01:05:08So I think again very exciting work
• 01:05:12would be great to to speak further.
• 01:05:15Because I think that we are at a
• 01:05:17stage and as you mentioned uniquely
• 01:05:20positioned within our department to
• 01:05:22make significant contributions to the
• 01:05:25understanding of how we might best
• 01:05:27advance psychiatric care for people.
• 01:05:29I couldn't agree more mark.
• 01:05:31And I think that your point there's many
• 01:05:34different ways that we'll all be in some
• 01:05:36family of general linear models, right.
• 01:05:39So and and people can approach
• 01:05:40this from various angles.
• 01:05:42I think that you know to your
• 01:05:44point it's going to be the data.
• 01:05:45That you guys have and we're going to
• 01:05:47continue to collect and it's going to
• 01:05:49be about the right behavioral response
• 01:05:51mapping which all of you are alluding to.
• 01:05:53So I I couldn't agree more and I think
• 01:05:56it's just it's it's just kind of you know.
• 01:05:59Maybe it's the sunny day,
• 01:06:00so I feel optimistic,
• 01:06:02but but I I I genuinely think that this
• 01:06:06was not possible 15 years ago, right?
• 01:06:08Like we didn't have the tech to do this
• 01:06:11and we actually now not only have the tech,
• 01:06:13but the information to do it.
• 01:06:14And so I want to just leave
• 01:06:16people with that idea,
• 01:06:17right, that that you know.
• 01:06:21And the fact that Yale is really
• 01:06:23stepping up into psychedelic
• 01:06:24medicine and and that you guys are
• 01:06:26doing this work and I couldn't
• 01:06:28be more supportive and whatever,
• 01:06:30whatever you need on on,
• 01:06:31happy to help.
• 01:06:36So with that, I think we
• 01:06:38should close for today. Alan.
• 01:06:39Thank you again for being here with us
• 01:06:43today and for covering this material.
• 01:06:46I believe next month, tentatively,
• 01:06:48Cyril has agreed to present either he
• 01:06:50or someone from his group or about some
• 01:06:52of the work that they've been doing,
• 01:06:54perhaps about DMT,
• 01:06:55where they've done some of the first work,
• 01:06:58both in health and in individuals
• 01:07:00with depressions. That'd be exciting.
• 01:07:01That's not confirmed yet,
• 01:07:02but we'll send out emails once we
• 01:07:06have a confirmation and a title
• 01:07:08and hope to see you all then.
• 01:07:10Take care, everybody.