Player is loading...
NIPMAP: Niche Phenotype Mapping of Multiplex Histology Data by Community Ecology.
Dr. Jean Hausser, Assistant Professor, Karolinska Institute, Sweden
Wednesday, Jan. 25, 2023 · 10:21 a.m. · 22m 04s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
so we've heard about the excitement around you know all the things you can do with
00:00:04
with the planning does this include in context of of of by region and in concert
00:00:08
as i was somewhat disclaimer this talk has no dipping running the different but
00:00:13
there's gonna be a i especially the kind of us an unsupervised a. i.
00:00:17
which we're gonna use here you know that to figure out how we can interpret a church architecture
00:00:22
because um the architecture of tissues on the lies there have a function and a classical example of this
00:00:28
is fun in the context of um the the lever
00:00:31
which is organised into hexagonal units of paper to sites
00:00:34
where around these units you have a arteries the bringing oxygen and nutrients in the material veins
00:00:40
and what happens now lever is that they have but this has or close to the arteries
00:00:44
perform metabolic expensive functions such as the calculations on the question is is
00:00:49
and in the centre of the of the motif you have cheap chip cheaper functions like a tree that's right and this is
00:00:55
another example of tissue architecture is found in the context of immunity if you look at the lymph node is it
00:01:00
be says only does it she says on and the b. says first need the antigen in the b. says on
00:01:06
but for you know the for them to initiate a par full antibody response against
00:01:11
pathogens the action need to travel to the t. says only need that he said
00:01:16
and so correspondingly the breakdown of tissue architecture a correlates with
00:01:19
disease so an example of this is fun in the pancreas
00:01:22
where they are anatomical structures known and as a i that's where you find out faster than better says
00:01:28
and was says are responsible for maintaining the right because some church concentration the blood and the
00:01:33
destruction of this architecture correlates with the onset of type one diabetes for example when he's at
00:01:38
his sister to invade those diets and finally and last but not least in the context of
00:01:43
cancer uh this architecture measures as well because humans can be stratified into inflamed and not inflamed
00:01:50
depending on the abundance but also on the spatial distribution of t. says
00:01:54
that you know and this distinction as a major importance for therapy because um
00:01:59
the inflamed humans respond much better to not r. p. compared to non infringements
00:02:04
so because of the relevance of just architectures uh in e. to buy the
00:02:08
gene having disease there's a lot of excitement around developing technologies to profile a
00:02:13
dish architecture at the fundamental units of tissue beta g. which is the single
00:02:17
said and so one such technologies could maybe stands for multiplex engaged finding imaging
00:02:24
and what maybe allows you to do is it i mean
00:02:27
the local abundance of uh dozens of protein markers within your tissue
00:02:31
um and so you would design the panel choose that thirty six protein markers
00:02:36
i mean you measure the local abundance using mass spectrometry and then they and then
00:02:39
i think the image you can determine ways are the positions that individual cessna tissue
00:02:44
and then use mark information in order to determine what is the type of it said the second to said he said
00:02:50
and also what is the phenotype of each said is this
00:02:53
a quiescent proliferating activated migrating and so on and so forth
00:02:58
now there's other technology to do the same another technology for example is i. m.
00:03:02
c. was so based on my sedimentary a way which um are yeah he's been
00:03:06
pioneering over the last three years especially in the context of cancer and that was
00:03:10
the technology is based on human infrared sense code x. um for i dislike if anymore
00:03:16
now what is that i i i was uh to do is produce um multiplex
00:03:21
images um it that that looks like this where this isn't the context of breast
00:03:27
cancer it's a tissue section from the cheaper negative breast rimmer and every dot represents
00:03:32
one cell and the stanza called according to the type he said he said kansas at
00:03:37
and what we can appreciate here is that there's clearly some kind of special organisation right there's a
00:03:41
low density zone of says on the left with is of being says which happened to be cancer says
00:03:46
yeah and then on the right you have green says which represents the
00:03:50
says and and t. helper says but what this plot doesn't show is that
00:03:55
for each of those says we also have information about the intensity of seventeen infinity
00:04:00
function markers what the size of doing and so no there's a representation challenge here is how can we possibly
00:04:08
represents this information on to these plots and i wanna post for seconded to to to to think about this 'cause it's it's not true
00:04:18
so to push it well this is a challenge you have to think what we try to achieve with this data so typically what we want to understand is
00:04:24
all this says that have a specific filler types and the and the spirit
00:04:28
of the specific spatial structure for example if you look at proliferating can say
00:04:32
is kinda says do you find any specific locations dreamer which is important because
00:04:36
then you know where to treat so answering this question is is difficult because
00:04:41
what you need to do to answer this question is you need because say type and then you need to pick
00:04:45
a phenotype and it's important to pick cell type first because if you pick the prefer eating phenotype in matters a lot
00:04:51
it is gonna type happens in kansas as earned union says in something different from
00:04:54
biology so you need to be given a type i said type and then you
00:05:00
can make a picture perhaps where one dot is a cell you only look at
00:05:03
kansas as and then you call this says according to how strong it expresses phenotype right
00:05:07
so you have seventeen personal definite type seventeen percent of cancers
00:05:10
said forty samples that's eleven thousand images for these data sets so
00:05:14
if you wanna go for there to find special structure there's gonna be a lot of work uh for for a scientifically speaking
00:05:20
but actually gets worse because most of the time when we don't care about
00:05:24
funding sums phenotype that has specific structuring space what we care about is really
00:05:29
special interactions between finer types things like whenever you have a
00:05:33
cancer cells with the specific uh like and then it's membrane
00:05:37
then the t. cell next to it happens to shut down right that's what we care about because if we understand this
00:05:43
then the this like under the kansas as can be a a target for therapy in order to
00:05:47
to make sure that it is says don't shut down when you're in this environment of kansas as
00:05:52
but now to represent this unit to pick a phenotype an egg and it said type
00:05:56
and you have to make this to square this because you have to like it was all possible interactions a few types
00:06:01
that produces eighty for thousands possible images for each section
00:06:05
that that you're looking at so going for this even if
00:06:08
you are very motivated scientists with that of timely hand like that's this sauce to the needs of what humans can do
00:06:13
so clearly there's a need for um more principled approaches to
00:06:16
summarise definitely big architecture of tissue and find interesting field types
00:06:21
so two approaches we can take advantage of biology of the
00:06:25
better just issues which happens to have a a local nature
00:06:28
in the sense that the the tissue but it emerges because cell as self organised by communicating
00:06:33
with each other with receptor like ends in a membrane that interact with the next to each other
00:06:39
or by making us thing is that the few is located in the tissue and influenza said behaviours of setup located close by
00:06:45
um so one example of this is the that under system that
00:06:48
establishes checkerboard patterns of of of phenotype so over over to shoes
00:06:53
um so that's locally nature proliferation stoker because the offices are located next
00:06:58
to each other and not this t. v. in random places and tissue
00:07:01
on migration is local because migration happens by local progressive movement for the
00:07:06
tissue and not by this the gems like a fee would go around in
00:07:10
in in space and immunity is is quite a bit of images look as
00:07:14
well 'cause uh it requires interaction between human set and and the pathogenic set
00:07:19
so it makes sense to start hours expression of dish architecture locally
00:07:23
and so only we can do this is dick specific places the tissue and then look
00:07:28
at what is this or composition there and if you do this uh uh you can start
00:07:32
seeing patterns for example maybe see there's a lot of size where there's a lot of
00:07:36
women says and there's gonna be another cluster of sites with that sort of kansas as inside
00:07:40
and so if you see this you know that i would suggest that this to initiating the tissue that structure the tissue and union ish and
00:07:47
a cancellation and so if you do to this for a lot of
00:07:50
sites in your colour decides according to the ammunition and the cancer cluster
00:07:54
then you can automatically segmented tissue into dishes and having this you
00:07:58
can then ask all the specific set added associate with a certain age
00:08:02
other specific filler times though current business to don't occur in the knowledge and this id
00:08:08
had this idea has been the basis for interpreting multiplex sister to that uh so far
00:08:12
and it's proven extraordinarily powerful x. interpreting to show architecture in the
00:08:17
context of health and disease now this works really well as long
00:08:22
as the tissue as a relatively simple structure relative
00:08:25
to that sites that you're studying because in this regime
00:08:29
a site is going to be locating meaning in each and therefore a set contains either immune says or
00:08:35
cancel said so you get clusters but now look at what happens when the tissue as a more concretely structure
00:08:41
where there's a lot of interface regions between the the cancer in the new niche in
00:08:45
this case the sides of finland its interface regions and now you don't get clusters anymore because
00:08:51
the seller completion of this site that so the interface between the cancer cells and and and diminish
00:08:56
is going to be um a mix of of what you find in the cancellation and in the
00:09:01
munitions is gonna be a weighted average of those two dishes depending on the exact position on the interface
00:09:07
and when you look at that aside what emerges is not clusters instead you have it in your continuum
00:09:11
where in the end of the lines you get size there are located within a specific niche
00:09:17
like in the core finish and in the middle of the line you get side that are like it is
00:09:21
somewhere at the interfaces between between two dishes so in this case it's not a good idea to clustering because
00:09:28
first of all you you made a lot of close sort of capture line or it's not clear
00:09:32
how many classes when it enters the sky's the limit right you can have one closer to side
00:09:36
and that is a real danger that we haven't inflation of clusters that may blur due to
00:09:40
should better g. great 'cause we get lost in the number of clusters it can be uh
00:09:44
strong cognitive burden like if we can keep things it things simple we should keep in simple
00:09:49
um another problem is that in clustering there's no notion of hierarchy between what finish and what
00:09:54
an interface and so you know the to address this we get inspired by ideas from community ecology
00:10:01
because it turns out that commutes ecology tries to sort of a similar problem to the one we wanna sort intuition biology namely
00:10:07
they want to understand how individuals of different species go heartbeats in different encourage finishes
00:10:13
so the metaphor there is that in the video from different species or says uh different types
00:10:18
and different naked gun issues are so that you can dishes and it turns out that up until the nineteen fifties
00:10:24
comedic urges would use the same analytic framework that we currently use to interpret multiplex sisters data
00:10:29
maybe they will cluster size that would just a species then in the
00:10:32
fifties good looking around and what is suggested is to do something different
00:10:37
he suggested to represent sites on x. is that represents
00:10:41
the the a species of the sailor composition of the sites
00:10:44
and look for structure there directly um and what you can do there and
00:10:48
then you you can check is there across the structure to it and if not
00:10:52
this kind of plus can tell you what could be better structure to interpret the architecture of your eco system or a beautician
00:10:58
so applying this id now to tissue but urges difficult because
00:11:03
this works well when there's two species or to set types right
00:11:06
but in this data set theory have seventeen setups so the what this means is that now we
00:11:12
we have like a seventeen dimensions space and every dot represents a side of the specifics or composition
00:11:17
we can't navigate this infrastructure does in this as humans
00:11:20
because when for g. unfortunately was second free dimensions so
00:11:25
however we can use techniques from statistics such as a principal compress nine is is to reduce the
00:11:29
dimensionality of the data and so if you do this you get a plot that looks like this so
00:11:35
the axes all the first three principal components of the data and if you're not too
00:11:39
familiar with these techniques just i consider that every x. is he represents the seller composition
00:11:46
and so the exact says that are on the different axes when i can't find the minutes but for now just imagine these
00:11:51
are just axes of say a composition and so if to such a close to each other women that have the same server composition
00:11:58
and so here we have we took a hundred sampling sites for each of the
00:12:02
four teacher section of this data set so there's four thousand dots on this picture
00:12:06
and what you can appreciate is that is not that is not there isn't there isn't a clear cluster structure in these
00:12:11
data sets where this more like there's a it's me or continuum of of sites in terms of zero composition and moreover
00:12:18
those sites have a specific structure neither land on the three dimensional simplex apparently we're
00:12:23
trying our bases and that's interesting because it suggests a specific interpretation to detach architecture here
00:12:31
and by the way this uh structure of the simplex we we don't join by hand right you can if
00:12:36
this simplex refitted using algorithm from satellite imaging so this
00:12:39
is done automatically in an objective fashion by the computer so
00:12:44
if you say simplex it's important for interpreting tertiary architecture because
00:12:48
any point within the simplex can be on this to there's a weighted average of
00:12:52
the endpoints that defined the simplex so biologically the saner composition of any point in
00:12:57
the tissue can be understood as a mix of for his surgical dishes that are
00:13:01
identified is the endpoints of the simplex so if you see a side close to
00:13:06
the extremity of uh of the simplex in is that this site is updated the core damage
00:13:11
and if this site is somewhere in in between the two endpoints in in the decide come some interface region
00:13:17
now we can use this idea to automatically segments tissues into niches by completing the
00:13:22
weight of the different dishes a different position that issue and calling the tissue accordingly
00:13:28
we can also use this id another to automatically find interface regions between between
00:13:32
dishes just by looking provisions where to unleash have high importance not just one ish
00:13:38
and then the result of businesses looks like this so here we looked for a uh the
00:13:42
interface region between inflammatory in the in the cancer niche and this is advancing the film because current
00:13:48
method is um that exist to find interfaces i can be difficult to implement because they have
00:13:52
different parameters we need to set an extra for this to work and this is fully automate sized
00:13:57
so it can be helpful to find inflammatory it
00:13:59
cancer interfaces which is our central important fortunate biology now
00:14:05
we're talking about but that button issues so they start to get really and uh to and to explain what businesses are
00:14:11
so to interpret the meaning of the dishes we can look at what is this error composition so if you do
00:14:16
this you can find that for example to bluish here is there's a lot of kansas says it's the cans image
00:14:21
and you can do this kind of reasoning to explain
00:14:24
that the um brackish happens to be the fabric picnic recognition
00:14:28
and then there's there's an inflammatory niche with that of t. says and and uh uh inflammatory macrophages
00:14:34
and there's a pinkish whether it be seventy upper says it suggests a tertiary infrastructure niche which
00:14:40
has been proposed to serve in in mediating antibody response pathogens so
00:14:46
what's important about this is that those dishes or not you right if you ask about that
00:14:50
are just it would tell you well not surprise we known about this for ten or twenty years
00:14:54
uh and that's a good thing because it should just dates this um
00:14:58
i've been issue approach can we discover things that we already know solidly
00:15:02
and then the question is can use to discover novelty right 'cause this is new
00:15:05
technology it's expensive to generate is that i can find something new into that up
00:15:09
maybe i'm not the question could be should be find something new or is this enough to
00:15:13
explain to show architecture so to find out what we've done is to continue adding one issues
00:15:19
and what we find is that information is already capture eighty two percent of the spatial variation deter composition
00:15:25
of shimmers so it's already very good summary and as we add more in one ish is it's not like
00:15:29
there's a significant increase in how much of the church architecture we can explain so we suggested for is probably
00:15:35
a good place to stop in addition the extra dishes don't have new biology is just more of the same
00:15:40
um another consideration to make is that we we use the clustering approach
00:15:44
should try to see if we could find interesting munitions in the data
00:15:47
and there for so what will deserve that if you use for classes you explain less
00:15:51
then then if use for dishes and that makes sense because clustering is a general purpose approach
00:15:56
whether danish is that come from ecology or tailored for to show architecture so
00:16:00
it makes sense that they are able to capture that they are better but
00:16:04
even increasing the number of clusters we never explain more then then phone issues
00:16:07
we explain so it suggested for dishes is a good place to start in addition
00:16:13
it turns out that those furnishes also able to explain inter presentation
00:16:16
variation in the cellar composition of tremors and that's a applies across
00:16:22
um because there's one issues come from people negative press streamers and we can with this you
00:16:27
can explain the interdiction variation in hundred and fifty tremors that come from different breast cancer cell types
00:16:33
a network profile with a different technology on a different continents uh and that are actually um
00:16:39
for those of magnitude larger then the data used to find inches in the first place so what
00:16:44
i'm trying to do here is to to convince you that the um this kinetic energy approach
00:16:49
is able twenty finishes that could provide an objective
00:16:52
quantitative a foundation to interpreting to show architecture so
00:16:57
how can you know you this you how can we use this to uh to answer the question we initially set to answer namely
00:17:03
how do we find a uh how can summarise affinity be architecture of tissues and how do
00:17:07
we find interesting thing that acts well with this dishes now attend what we can do is
00:17:13
we can select a specific set type and asked if expression
00:17:17
of a specific phenotype correlates with okay station in a certain ish
00:17:21
and we can do this by using a um statistical measures of
00:17:25
of associations like this german correlation uh a coefficient and then we can
00:17:30
iterate over always say types all phenotype sin own issues so that
00:17:34
you can save the work of doing this by i for the human
00:17:37
and if you do this for phenotype said types initially is it the result looks like this you get a hit not that tells you
00:17:44
which said type have which phenotype in what niche i mean what
00:17:48
interfaces so the way to interpret this not for example is that
00:17:53
if you look at b. says you see that be says express
00:17:56
the ha lady are marker in the a. t. at its niche
00:18:00
so this is still quite a lot of data uh and we can further
00:18:03
compress it into a tissue architecture table that looks like this where for each nation
00:18:07
you indicate oh which uh phenotype in which state types or representative of the snitch
00:18:13
now um what i wanna do to finish his presentation you show if
00:18:17
examples of interesting thing the types that uh that that the come out
00:18:21
of this analysis so one example we just talked about right i is
00:18:26
the says so what we've done here is to take a representative section
00:18:30
from the data set and every daughter presents and besides we removed all
00:18:34
other states here in other to clarify the plot so it's there's more
00:18:37
says that we don't show them so going you could be says the pieces of coloured according to the expression of the h. e. d. or marker
00:18:44
and then the colour in the back one represents the new segmentation
00:18:47
that that we performed automatically using the commute ecology uh approach dimensions
00:18:51
what you see it going you push it here is that the b. says express the image because to marker whenever
00:18:56
they're in the green chili snitch might um this is classical in energy makes sense it's good that we can confirm that
00:19:03
similarly we we we discover that's a kansas says that express image just to our funding didn't matter in age
00:19:09
and that's kind of today's express emission 'cause one or found that the interface between the internet
00:19:14
region and the cancer region so nothing you here um but you can add to this approach
00:19:19
in other to discover new phenotype that haven't been proposed yet for example when we look at
00:19:24
macrophages with find that localisation within inflammatory niche correlates with expression of c. d. forty five are all
00:19:31
no that's interesting because c. d. forty five are what happens to be a marker of inactivation but mean in the context of t. cell biology
00:19:37
so here we see this macrophages would could indicate that c. d. forty five hour is a more general market of inactivation
00:19:44
another interesting observation is that we see that then with excess located in inflammatory region have the carotene
00:19:49
six marker so that's weird because then we think says don't make writing six it's in the particular marker
00:19:55
but this could make sense on the type of this is that the entity says uh trafficking for the cancer area and
00:20:01
then they come into new damage a revision and they carry over the cut in six marker over to dinner much religion
00:20:07
so you can generate a habit this is about union said trafficking for that your mike environment in this way
00:20:12
and finally we see that new trophies that express the p. d. a.
00:20:15
one uh like and look at at at the interface between the inflammatory region
00:20:20
and the cancer region yeah that's interesting because p. d. l. one is often targeted in in the
00:20:25
cherokee as it's a known to be an off switch for the anti constructed that your t. cells
00:20:30
so seeing this could suggest that interferes contributes to a helping kansas as evade immune system
00:20:36
so we've shown so for that you need not which is the approach to describe here
00:20:40
um he works well on on turner sections but actually the so works well on heavy sections in this case in uh
00:20:47
a a developmental done and we've illustrated so foreign protein based special profiling but
00:20:52
you can also use it with our new data so our summarised now by
00:20:58
recording that's multiplex histology it is they can begin changer in
00:21:02
interpreting tissue architecture in the context of head i mean this is
00:21:05
but the data is challenging and complex to to interpret and so to help with this we can use ideas from
00:21:10
commute ecology you know that to find what are the niches that structure the tissue and their interfaces and on this
00:21:17
nation interface structure we can project finger types in other to first
00:21:21
of all summarise the architecture divinity big architecture the tissue and second
00:21:26
in a identify interesting feature types to just generate but you could have but it's it's and so if
00:21:33
you find that you future somebody's approach you know the two unless you roll multiplex sister to that ah
00:21:38
uh please come in get in touch with me because we developed an r. and
00:21:41
tighten package you know the to do this automatically and uh we looking for beta testers
00:21:46
so i wanna thank any so who did most of the analysis on
00:21:50
this project is what is five you is a jihad yeah david veteran antony
00:21:54
also much nissan then searches set us and yet karen for for collaboration than
Conference Program
The Idiap Research Institute in Martigny is launching a new public series of symposiums
Lonneke, Idiap Research Institute
Jan. 25, 2023 · 9:39 a.m.
420 views
Interpretable artificial intelligence for cancer personalized medicine
Dr. Maria Rodriguez Martinez, Group Leader, IBM
Jan. 25, 2023 · 9:48 a.m.
5 minutes questions: Interpretable artificial intelligence for cancer personalized medicine
Dr. Maria Rodriguez Martinez, Group Leader, IBM
Jan. 25, 2023 · 10:10 a.m.
NIPMAP: Niche Phenotype Mapping of Multiplex Histology Data by Community Ecology.
Dr. Jean Hausser, Assistant Professor, Karolinska Institute, Sweden
Jan. 25, 2023 · 10:21 a.m.
133 views
5 minutes questions: NIPMAP: Niche Phenotype Mapping of Multiplex Histology Data by Community Ecology.
Dr. Jean Hausser, Assistant Professor, Karolinska Institute, Sweden
Jan. 25, 2023 · 10:43 a.m.
Data Science for Precision Oncology
Prof. Olivier Michielin, Head of the Center of Precision Oncology, CHUV, and Group Leader at the Swiss Institute of Bioinformatics
Jan. 25, 2023 · 11:24 a.m.
5 minutest questions: Data Science for Precision Oncology
Prof. Olivier Michielin, Head of the Center of Precision Oncology, CHUV, and Group Leader at the Swiss Institute of Bioinformatics
Jan. 25, 2023 · 11:54 a.m.
Applications of AI in oncology drug discovery
Dr. Slavica Dimitrieva, Associate Director & Senior Principal Scientist, Oncology Data Science, Novartis Institutes for BioMedical Research
Jan. 25, 2023 · 12:07 p.m.
5 minutes questions: Applications of AI in oncology drug discovery
Dr. Slavica Dimitrieva, Associate Director & Senior Principal Scientist, Oncology Data Science, Novartis Institutes for BioMedical Research
Jan. 25, 2023 · 12:32 p.m.
Recommended talks
Clinical Natural Language Processing
Carsten Eickhoff, Assistant professor of medical and computer science at Brown University
Aug. 20, 2019 · 11:05 a.m.
299 views