Computational models of stem cell decisions

13 views

Formal Metadata

Title
Computational models of stem cell decisions
Title of Series
Author
Marr, Carsten
License
CC Attribution 3.0 Germany:
You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor.
DOI
Publisher
Weierstraß-Institut für Angewandte Analysis und Stochastik (WIAS), Leibniz-Institut für Atmosphärenphysik (IAP)
Release Date
2019
Language
English

Content Metadata

Subject Area
Abstract
Molecular regulation of cell fate decisions underlies health and disease. In my talk, I will present mathematical and statistical models that describe molecular interactions, differentiation decisions, and single cell gene expression. We use these models to infer molecular and cellular properties from biological and biomedical data. For example, in lineage trees of differentiating blood stem cells, we often observe correlated state changes between related cells. Using these correlations and a stochastic model of the differentiation process, we find differentiation events to happen much earlier than previously anticipated. To predict differentiation prospectively, we use a deep neural network trained on image patches from brightfield microscopy and cellular movement. Surprisingly, we can detect lineage choice in blood stem cells up to three generations before conventional molecular lineage markers are observable. Finally, I will present a method for fitting stochastic models to lineage trees. Using a Bayesian inference method, we compare possible models of autoregulation, an important gene regulatory motif in stem cells.>
Loading...
Thermodynamisches System Decision theory Forschungszentrum Rossendorf Mathematical model Event horizon Theory
Forschungszentrum Rossendorf Moment (mathematics) Right angle Subtraction
Beta function Dynamical system Decision theory Moment (mathematics) Median Data analysis Mathematical model Mathematical analysis Incidence algebra Mereology Local Group Mach's principle Mathematics Scalar field Physics Units of measurement Associative property Units of measurement
1 (number) Sampling (statistics) Mathematical model Mathematical model Price index Subtraction Mach's principle
Fluid statics Beta function Scaling (geometry) Fluid statics Function (mathematics) Right angle Mathematical model Figurate number Block (periodic table) Greatest element Grothendieck topology
Slide rule Group action Divisor Interior (topology) Mathematical model Division (mathematics) Line (geometry) Mathematical model Translation (relic) Food energy Autoregressive conditional heteroskedasticity Right angle Module (mathematics) Uniform space
Dynamical system Concentric Element (mathematics) Mathematical model Mathematical model Measurement Number Inclusion map Mathematics Order (biology) Equation Equation Aerodynamics Gamma function
Area Complex (psychology) Mathematical model Aerodynamics Mathematical model Atomic nucleus Physical system
Mathematical model Right angle Mathematical model Hand fan
Right angle Set (mathematics) Number Physical system
Product (category theory) Term (mathematics) Charge carrier Energy level Right angle Division (mathematics) Mereology Functional (mathematics) Vector potential
Mathematics Term (mathematics) Decision theory Expression Right angle Stochastic kernel estimation Student's t-test Limit (category theory) Subtraction Symmetric matrix Physical system
State of matter Decision theory Direction (geometry) Right angle Parameter (computer programming) Mathematical model Physical system
Point (geometry) Number Theory of relativity Stochastic Multiplication sign Direction (geometry) Right angle Open set Descriptive statistics
Stochastic process State of matter Equation Right angle
Decision theory Multiplication sign Noise Mathematical model Subtraction Physical system
Multiplication sign Right angle Mathematical model Mach's principle
Point (geometry) Charge carrier Regular graph Curve fitting Mach's principle
Multiplication sign Quantification Quantification Right angle Energy level Set (mathematics) Student's t-test Mereology Number
Trail Quantification Set (mathematics) Figurate number Physical system
Point (geometry) Group action Term (mathematics) Multiplication sign Right angle Mach's principle
Point (geometry) Axiom of choice Process (computing) Decision theory Multiplication sign Decision theory Element (mathematics) Mathematical model Cartesian coordinate system Mereology Measurement Inference Many-sorted logic Network topology Noise Right angle Lipschitz continuity Figurate number
Point (geometry) Decision theory State of matter Decision theory Multiplication sign Algebraic structure Correlation and dependence Division (mathematics) Event horizon Wave Event horizon Analogy Network topology Right angle
Point (geometry) Process (computing) Process (computing) State of matter Forcing (mathematics) Multiplication sign Decision theory Correlation and dependence Mathematical model Functional (mathematics) Estimator Event horizon Function (mathematics) Network topology
Point (geometry) Process (computing) Stochastic Process (computing) Function (mathematics) Network topology Mathematical model Limit (category theory) Parameter (computer programming) Limit (category theory) Physical system
Statistical hypothesis testing Group action Inference Parameter (computer programming) Mathematical model Parameter (computer programming)
Musical ensemble Link (knot theory) Decision theory Parameter (computer programming) Mathematical model Parameter (computer programming) Set (mathematics) Likelihood function Number Group representation Summation Cross-correlation Network topology Inference Length Subtraction
Point (geometry) Trail Process (computing) Network topology Real number Multiplication sign Right angle Parameter (computer programming) Mereology Food energy Subtraction
Dynamical system Prediction Generating set of a group Decision theory Multiplication sign Division (mathematics) Right angle Aerodynamics
Point (geometry) Area Product (category theory) Concentric Multiplication sign Algebraic structure Division (mathematics) Differentiable manifold Line (geometry) Estimator Cross-correlation Network topology Aerodynamics Figurate number
Point (geometry) Product (category theory) Linear regression Decision theory Direction (geometry) Multiplication sign Mathematical model Number Prediction Symmetry (physics) Network topology Aerodynamics Relief
Point (geometry) Product (category theory) Divisor State of matter Multiplication sign Chemical equation Decision theory Division (mathematics) Mathematical model Student's t-test Prediction Linear subspace Network topology Right angle Aerodynamics
Point (geometry) Dynamical system Prediction Axiom of choice Multiplication sign Decision theory Aerodynamics Mathematical model Mereology
Prediction Multiplication sign Sheaf (mathematics) Right angle Mereology Mathematical morphology Equivalence relation Mathematical morphology Position operator Resultant
Point (geometry) Multiplication sign Image resolution Right angle Division (mathematics) Mathematical morphology Number Mach's principle
Convolution Statistical hypothesis testing Statistics Mathematical model Mathematical model Set (mathematics) Convolution Wave packet Mach's principle Social class Right angle Subtraction Social class
Summation Order (biology) IRIS-T Right angle Subtraction Mach's principle
Convolution Process (computing) Order (biology) Drop (liquid) Object (grammar) Mathematical optimization Convolution Units of measurement Mach's principle
Convolution Content (media) Right angle Mereology Convolution
Convolution Statistical hypothesis testing Trail Multiplication sign Mereology Perspective (visual) Convolution Maxima and minima Sign (mathematics) Maxima and minima Function (mathematics) Order (biology) Right angle Identical particles Friction Units of measurement Linear map Social class
Food energy
Statistical hypothesis testing Graph coloring Network topology Multiplication sign Order (biology) Distance Bounded variation Subtraction Wave packet
Area Summation Curve Iteration Prediction Equivalence relation Measurement Mortality rate Position operator Units of measurement
Trail Network topology Expression Division (mathematics) Prediction Mach's principle
Stochastic process Group action Divisor Multiplication sign Expression Trajectory Parameter (computer programming) Equivalence relation Matrix (mathematics) Many-sorted logic Lattice (order) Kinematics Phase transition Right angle Divisor Figurate number Social class
Wage labour Adaptive behavior Sampling (statistics) Counting Mereology Cartesian coordinate system Sinc function Social class
Sampling (statistics) Right angle Insertion loss Subtraction Mach's principle
Sign (mathematics) Prediction State of matter Atomic number Sheaf (mathematics) Hierarchy Set (mathematics) Subtraction Social class Mach's principle
Point (geometry) Statistical hypothesis testing Inference Prediction Algebraic structure Set (mathematics) Focus (optics) Subtraction Grothendieck topology Number
Shooting method Axiom of choice Consistency Moment (mathematics) Mathematical model Mathematical model Deutscher Filmpreis Local Group Number
Limit (category theory)
ch I it's going to be at the foot imitation i'll redraw and the mixed but a mixed crowd of people here during the event but we talk about
the brain research and also in the defeat of the sun they separate so I really enjoy these this a mix of public some people a to to be a all right I'm
from Munich so this is where I work it's center on the Newark Of for reasons the reasons for the for environment and hence when I give a talk somewhere I sometimes stop by saying this is due to the very north of theories takes you quite a while to go there and then and sometimes they as long as going to the place where I is is so
your true here to the hippocampus it took me really quite a lot have come here so the coming here if the travel did you should come to new of course so I wanted to go there and have a look so that there was and and it consists of roughly 30 different research entities it's quite diverse from groundwater ecology to spent research to accomplish apology what we do we have
about roughly 2 thousand employees 700 scientists Muslim PhD's that topic
over there is the individual being dominated by its genes but its lifestyle and at that moment right so that's how the topic of or that the same people can can
agree I get over there and we're part of the latter has much association with 21 centers all across Germany against Micro that so
great to be here great overlook what at the lab is as it is doing and I'm really keen on discussing with so this
from incentive-compatible and here we
are roughly 60 employees at the moment kind of growing rift and from research
groups and my research group is concerned with the dynamics of singer says and that's 1 good story
by the way we just started as a whole will start no later this year something which would haiku the head what's Association unit of 4 the once Artificial Intelligence Coordination Unit and we try to to spread out and AI methods to the whole and was community and we started also science school aquatic cooperate assigned to recruit more people and doing that kind of incident which
now we're talk starts what we do as all of you do right it's affection method be abstract complex systems 2 things the condition understand and of course this will become now in biology and I just continue with what the 2 because before me and I introduced for the couple of different ones symbology for for sample that might be different set of
1 solve this is a gene network every
gene and they're connected for season of every every thought indices and 2 diseases are connected if a gene is associated with these diseases this kind of model obviously on
quantity of right the sudden makes it's static and it's not scale because it covers all the diseases we know we do have other models like that and actually for
talked about is they call this a model just drawing 2 proteins that that's a model for biologists and you see to to to to proteins here pure wanna get a 1 and the model is was what would you condense is all the experiments you did before right it's the last the last figure in a paper and you say OK these 2 proteins they can bind to each other x 2 particular protein side and then they can do something at the beginning he beat the transcription of another all right so this model that is also quantitative obviously is also static but it's was scared just considers to and OK that just as the
short the tightest you know that but in a talk about the script effect as and members of the because home your also
divide you know there are genes they are
transcribed to messages and also translated to proteins and that's of gene
genes are made of DNA messages made of right in the producer made of amino acids I guess some of these proteins factors cannot bind to the genome and regulate the transcription of other genes let's
that so as a final model example no this is something that we connotation about just would call a more like a slide so we have a rather complex system with 10 players maybe you see so the gray line is the the outer rim of affairs and you the receptor and things can bind to this receptor and septic signal downstream into the cell and then things happened there energy inside dividend deactivated
and this kind of idea we cost into models for some an order of differently equation they say the change of this is the change of the concentration of a receptor on the cell membrane for example and has a couple of elements that determine that
right and then you can fit even this kind of to to and measurements and that's coming close to work or that's basically what we call a collective knowledge and you cannot predict things and then this kind of mollusk quantities if Reddy can write down numbers it's dynamic and it's the same measles gay because you can come up to 10
speakers of 2 of the 10 players with all right so that some of them will be over these different
models if you don't wanna just they also called other things models as I told you write little drawings but they also Co species models so for somebody call like is a species for biologists and it's a very simple bacterium is no nuclear fried so it's it's you can do all kind of genome annotations with it and a model because it's much simpler than the more complex mammalian cells on same with yeast now this has a nucleus but the you can genetically modify as much as you want and you can do experiments with than 1 system for may become more complicated animals like the fish than your areas of office also more lot
them it's great because it's translucent you can watch it grow in the in the in development
and of course the mouse is mated in all placing in unique and I for for drug research as well this is the most widely used model organism right because it's a manner but the you can do and words with OK so how much of what a good start what it just this pretty good the frame right where where our models of what you want so
and then talk about them by the 2 models knowing the more computation fans and I'm going to
talk about a singer says so this number is amazing right of 30 trillion
human says every 1 of us is consistent with this number of that is thinking and not just this number is amazing but it's also interesting to see that most of these sets consist of about society and I mean this would affect are most red blood cells the the the the tiny guys without too funny for blood right they transfer oxygen and most of all says these guys and then that's amazing number but even more amazing I think is that every day each of us produces 500 billion of these days are the sets right so we are not a static system but we kind of a dynamic approved a career right to introducing ourselves and we think know how is that done was the pro this quote him
out we is is the production of blood and
after hundreds of research and that's basically the dock what people believe employment with so the
ideas you have with extensive and these guys we roughly estimate that we have between 20 thousand 1 million nobody's and they sit in Albemarle and they don't do much but this this sometimes divide and then they give rise to a more differentiated so for example short term about Christensen and then going down here at 2 what's the functional says right but do we really things the potential of the says decreases right so if you are a carrier Cyprus as a producer you will be able to to to become better so but you are very close to becoming a red parts of which 6 doing doing stuff and in the body right so you lose potency and you go toward a function of said going from the stand said over the progenitor cell level to the function of
assuming a bit that's 1 thing now we looked into a bit and then to use that so many in the in the even more so you if you a common myeloid progenitor my because these guys 16 Road and this so tight is believed to have to produce 2 different potential limits which are more mature right either they occur side there was antigens of left and kernels side also approaches of of OK so what do we know what the system well if you if you kick out this wanted the damaged earlier right Pugh 1 in mice of course because their candidates from its if you kick up to 1 in the gene optimized you won't get any duties if you keep out gather 1 you won't get any any peace right so from these experiments can conclude that the student which affect us to 1 get 1 of this are probably very important for the for this difference decision way and then if you think about how is that such a sea of peace and able to make this decision either to become a and B boards to become a G. then the student would affect the proper important
and now this again what I showed earlier right Our from some other experiments people figured out that if you wanna get 1 confined to each other and then inhibit the expression of the other so that app ideas together with the French the importance of these 2 produced visit conversed an idea that Q 1 and and 1 are former Thomas which OK and that's what I'm I'm sure that over there so got 1 is inhibiting Q 1 Q 1 is inhibiting got 1 than they might set at the rate of of and is now this has been done to toggle switch right and you can think of it as baby show that so if 1 gets a bit of an advantage it inhibits the other 1 and chapter down and then you go to want what 1 lineage or the other 1 is a bit of an advantage and then it should start in the morning to the other which way and this summer's which organize the model and so I can write down and ODU where you say the change of the 1 protein is determined by a safe activation tone and then you have you should inhibition turn right so if the other guy down here the latter that is the less increase of X 1 you get and yours of decay term because these proteins are degraded over time and if you do that in a poverty symmetric metaphor the 2 proteins and you simulated you look
at the of salt solutions so the system you find that if you tune the parameters in the system appropriately you see that you have 3 stable states of the system 1 in the middle where both Q 1 and the other 1 I expressed and then to other states where either the 1 or the other is expressed and that's nice because it fits basically to
all about its tuition right all of the the idea we have the system you can start with a state in the middle where both states can read but then you to when 1 of the parameters and you either go to a bar to that alter the direction right of this model is a nice wait to abstract or to to to describe what we think how these progenitors differentiated become more functional in such a to physician decision right
and now in 2000 though i in probably 10 years ago what a rotten co-workers shows that RNA is actually transcribed in bursts that means that the the DNA with the which code for the gene right this goes of most of the time but they sometimes it opens and then many polymerases 1 with this was changed direction produce many em nowadays at 1 time right so you get a lot of on a switch to the burst and then the gene shuts down again as coast and to describe that actually is not really appropriate right for that year rather need a stochastic description of the system and so we
came up with animosities at that point and tried to to just cast this told us which into a statistical relations and then you have to write
down the equations for the stochastic system this is complicated to solve it the only thing you can do with stimulating and that we do and if you do that you
you can also simulate a typical said objectively and then you find it if you start in the middle and underside state you see how such a simulation gives you the 1 or the other state and actually that fact affections you a bit more complicated as compared to the to the be case but you did the same thing right starting from 1 underside state you either go to the 1 or the other search all
right well we did also another model where we consider more players but then we have 11 players and become 1 that's too has to do so we used a Boolean framework here and it's not just as many people in particle physicists like this idea of the topics which is a model for a different vision decisions and many people and simulated that because it's such a nice system there's a biological relevance to it and you can still simulating understand and what's happening over the years when more these models appeared it became clear that well we can model it right we can we we haven't actually proven that is usually vision of Q 1 a got 1 is really decisive for this difference Asian decision in blood cells so actually this discussion a troubled our our
collaborator at that time conferred noise at ETH and he had an nice system to actually look into the system and that's using Panitz microscopy so taking which is much grimaces every couple minutes and then you also have a now was 2nd
model system right a mouse where I'm at that you wanna look within the gene known that a yellow for frozen simple protein and the got a 1 locus who headed the reference in protein knocked into the G 2 so whenever this gene is transcribed this attached to this protein which now lights up in yellow red so you can watch the protein and transcription over time so
what you mandate is breeding mice with these constructs here sorting out then of instances from the mice putting them under the microscope and doing kind movies right because idea was well let's watch T cells differentiate and then you can figure out if you wanna get a 1 are really doing so that's the
data the bride freely which how they look like we take images and every 5 minutes and you've seen our Hollis's run around these the themselves and some point they grow would differentiate and
we not just have despite feeling which is a toad regular information about the yellow for protein so this this tactic you 1 so whenever you see yellow shining up this is a mark of hope you 1 being expressed and then we also have got a 1 and as a marker so we know how to how much gas was invented to serious but we also have CD 16 32 or this is another marker and whenever we see this coming up and so we know that this cell has adopted the Gymnich you can't turn it back from there and then we also see this latter says for the major carriers side and you see that here and that's true 1 is negative but you get a 1 positive and 16 that 2 of them so that's basically the dataset we have bright fit images every 10 minutes then every half an hour we do this growth and images and now you're trying
to that question if you wanna get 1 or incisive all idea was to to quantify protein that isn't in a sense but so you know in 1 particular said how much you want on which kind in there and when I watch that over time so we need to trick the which follows in the sense of what and now to to to to come up with solutions for that you have to do what is now called file image informatics like here this data set which is a movie in just a speck of images and to distilled numbers all of that give to a to apply a cup of 2 was and if they're not around the right tools have to develop them and actually we thought that it might be easy to get the numbers of of that but it turned out that this 2nd the difficult part of it and it took us a couple of years ago the 50 students to have to come up with a solution for that and I cut that part
short I guess if you're interested in just a thought after was so we came up with the solution for segmenting the singer says in the movie flight and find the thing is that in there then
the road and inefficient soft but would also to track the says in the morning and that's actually not done by the computer but automatically so as to the system follows the same users overtime
and this knowledge this this this was for that of a conic vacation so whenever you have for track the says so we again with segmentation of the cell and which gives you this mask here and then we just sum up all the intensities in that mask we also releasing a tool that allows it to efficiently go into the data set and figure out where the segmentation favors and adapted and then finally in all these movies with a problem
because you know the images we have more photons have been generated in the middle of of of of the which for one-photon you put in as compared to the to the edges of the Richard so it's an uneven information and to correct for this setting effect and we came up with a tool that estimates the background some of these images and using the assumption that the country be the same in all these images if if we do that we can make which is flat and compare says all of which putting mother together we were at the end able to attract the
science the right and then when this once the cell divides you trick the Sisters of the term and then again this is this and at some point CD 16 32 is coming up in the 1st 10 of these images and that the guy who checking that takes that and says OK now the red cell from this time point on we know that this year all right and so what
we got after applying what these tools to other days this kind of trees right no time on the X axis and you see that there's a mother dividing sisters sisters and songs also at some point you might lose the cell because it's running out of the of the image some sometimes might just die and then at some point you are not able to trick any further because it's so dense and you can't figure out from 1 time for the next which sort of which but in heading that we also have information about how
much you wanna get a 1 is in The says well for example not only once a year and in that says elements was rising toward its rising than some point the cell divides and at half the amount of tuning the 1 in the sisters' and then this rising again and so on so long when sound we did that for like roughly 80 trees in the 4 experiments and we had these trees and then again what father this question right if you wanna go on decisive for the lips choice and if you look at that trees even many of them and for a long time it you realize that just looking at the 3 thousand had joined question because I don't really know where to look at right the data structures complicated because you have these trees and give a lot of noise in it in the measurements but maybe also in the in the process trees tryptophan is it's as a you and then you you also have the societal right that always the proteins half it at each time the cell divides the throat just by looking at the trees it's it's really not legend of not able to atOnce Krasnogorsk so the whole idea was to have from the 3 the out when the decision for a particular different Asian step occurs and then once we know where it happens we can look into that time point and compared with the mission vision will figure out if that morning future OK so the 2nd part of this talk is to code retrospective inference of french fries because from the trees will figure out where this debt now how can these be that done and the news
analogy so this is time this is my group into light with 15 that guy over the year over the next these she further the guy who who implemented that stuff on would show you she which were posted to Seattle and now if I would see him in Seattle so a in the part of of of them says decisions you could say he's making it a state change right from the ICB where about before to ISP wave excel at so if you would want to see him at NICB in 2017 right you would not be able to figure out when he made the decision to go right could have been to is a go for the we don't know if me he would
divide at the says anomalies to right and you would see in and in you would have a decision before the division of him and you would expect that she shows up in Seattle and according to time points where it might be a bit because 4th in middle but he would show up at the same time what if the decision after the division by then you uncorrelated time points often popping up a so that is to use a tree structure to figure out where the decision for particular event happens and of course in our case the idea comes from
looking at the trees because if you look at this market 56 232 which is the market for the 1 particularly it using that the is coming up in a kind of protection or is not all of the places that somewhere there somewhere so it's mostly happening here in the force of estimation and that's a pet everything all the trees so we had the idea
is to model this different Asian process and we basically have to step in here the 1 is equal to point process by which a state at 1 time point the said just decides to differentiate and then this problem is probably due to different might be constant all we'll we allow it to be a function of time in the function and then once the deceit the decision has
been made and we say we start aristocratic she's Persian process and and then at some point we hit the mark that the limit the marker of where the market can be detected in the movies and and that's basically the expression of off the market if we think it's grows so
unofficial as it here with the trees so this said he decides to differentiate and then this process starts suppressor genes which process and some point you hit the detection limit and you say OK here this market and the 16th of 2 marker and then the celibate later and this happens in the system as well and again it takes a while so you have a different process and you have a delay process and these 2 things a parametrized with them of parameters in all
model know with this model and revoke this parameters in there and this college the day and 9 1 and figure out what the parameters are so this 1st test if we can figure them out in a simulated test
things get paid right so what we do is we Singletary's which look exactly of which look a lot like the data we have and if you see Willard we know of course when the transition happens and how long the delays in when the market comes up and then we forget what a mark of all this delay in just the market is with GM and now we won from a
number of using the trees can be used to back these parameters of the model and we tried
this and for that we need the likelihood of the likelihood of of serving a particular set of trees given these 2 parameters and the way set up and it's a sum over well 1st differences tree and there are many different possibilities houses how this can come apart right you can ever an early decision long delay or you can have a late decision with the short delay and the links of these things are determined by the promise and then from each of these hidden trees we now decomposing to come of subtrees which can use you model and music of more representation for the subtrees to figure out what's the most likely parameter for a given length of delay for example in correlation of this no we put that only in in like to function and data with it and cursed
beast assimilated this to make sure that that works we find that from simulated we can infer back the right parameters and then we apply the real data energy that's the data we use here so we see this all the trees that make this G image the 1 part of the challenges and you see that this is quite a genius and that comes just from the fact that the the biologists attractive they sometimes track the whole tree because abortion with that and there sometimes the only track 1 right but nice thing is that we can put it all into all our elected estimation and because also that you would give you some information about about how long the process so we put that in that assuming here
and that kind of those trees we observed and then we forgot for each of the trees what's what are the most likely the most likely parameters and then from that was the most likely time point of different Asian
and this I show you now here so the step from black to gray that's the type 1 B we believe the decision has been made and would you realize is that However them comes up with rich long delays right so from the temple of division to the mark on that it takes 5 to 6 generations in all these things now we can do now is look
in what's happened there in all the data right so we go a long time and we pick 1 of these branches and no we look at the true 1
dynamics of the what is to choose which affect as we check what's going on in there and that's how
again such a time course looks right so you see that she won intensity rises is half prices out and so on so on at some point the market comes up and this point over there this is all predicted the condition this is when our algorithm believed that the division has remained from the correlation structure in the trees right we divide the
intensity by the area to get a bit of a smooth curve here and then we fit each of fulfilling thing so we figures concentrations with a straight line which is roughly an estimate for that and for the production of 2 1 that's and we can plot
this slope of the linear fit reaches production of you 1 basically to be compared for the for the federal relief that the decision has been made with duration 0 recorded and the compare that to 1 direction before and moderation of I other compared to the and Rossini 16 32 it's definitely there and what you can see is that if you want corruption does not change at this time point where we think the decision is made and chief you look at the model you would expect that you you know with this topics which right in and Q 1 is winning look I 1 then at that time point to 1 this is should off right so from looking at it we know we felt where that's not what you would expect from the model but of course it's not not enough to reach the model found however we thought we can do we now want to reject the model is just
taking this topics which model implementing it them stimulating and raw if it is same number of of trees that an experiment and that's checking what will begin from it right so we symmetries here with
a promise been for and then from all these trees down we look again at this time points for everything that approach and happens that you in that model we don't know whether whether decision happens because his topics which thing right it's a it's a it's the delicate balance between the 2 factors at some point tears right but where accepted the decision happened in that imposes Fudan so we also infer for for those trees with leaves a decision happens and then we look again how
through 1 is on developing their and for the simulations to see that at the time point where our with them I think the division happens there you see indeed and significant increase offshore production from 11 which is on the to a to a higher state so in the model where that when the button students to 1 is high produced in the data we can see that and this leads us to the
conclusion that the true 1 dynamics me the things as is inconsistent with the machine vision often shots so I could say well this is a bit of this by the method for rejecting modern but where as if you know right that's all we
can do or that's the best we can do we can reject models and all this does not fit to the data we have and basically this would here OK and
understood part of this story is known questions to that actually interrupt me whenever 1 the future from aggressive OK so so this was the the the reconstruction of the decision time point now that we know that decision happens before the market comes up the question is can we we
predict the fate of fair for this market section visible right because we know that position should happen or fact extrusion before I maybe we can predict what what's happening and the idea is what what we have we have the morphology of the says and with the speed of the says because we watch them in this bright food at 3 minute intervals right so then we have to identify the sentence by the channel and then this applies machine learning again as mine this because for it it right because that might help us to predict what this I guess all problem is a bit easier than the new chemical problem with treatments and the final results so I just look at that the education so part with for them by by fuel and I should for what I get sees if
occasion in bright feed we adapted to bed and it works so we can identify the things over time and then once we have to be can cut them out and we get the hidden patches of
sciences and since we have many of the service and we have a high time resolution you have roughly 2 million of these patches and now you can think that to win and that the high number as strain machine and 1 with it that should work so supervised machine and if I can do
that so you see no you can order them along the checkpoint right in full of says in bright future and you see that they are a little bit more they ship beginning in the role of it and as you see if you could so size that the cell will divide growth light crude right and again again at some point now they turn on the C 6 into and using this markets up so at that time here we know that this is a GMM and because the sense annotated down and what was as before we know that they would turn on the mark at some point but they haven't yet so we call them later OK so was machine learning will have
modern we use convolution whenever for that and since this hasn't been introduced all right that I do that briefly so
I mean we all agree on is that their mathematical or mechanistic models and statistic models that assume so that the enhanced I should in that side right
so very short introductions was pushing the hope boy you so with machine learning at
before the images right work in that way that different classes in the training set and test set and you take an image to extract
features from that image and you use that features fortifications from for example you have this image
of of this flower here you click the pedals and then you calculate with a high for
example right and look at different flowers you see that they differ in these features so
if you look at pet with of the different than tosses of flowers you see that you can use that feature discriminate these committees station performance however is then highly dependent on the feature extraction so we have to be able to extract the features and then have to know which features to expect in order to make this difficult work now 7 years
ago the idea of convolution unit folks has has disrupted the feet of image of computer vision it's by the idea of making this feature extraction offered on obsolete but integrating it into your learning process so this conclusion that bugs they are based on the idea that and that with its hand learns the optimal
features that you need in order to to justify your problem a great so it's not longer you who who who segments the objects and its features but it's enough of it and how is that done so this
Commissioner networks have a cup of these layers convolutional layers 20th and so on and the
conclusion is basically our simple right you you just take an image and you just might apply and part of that image with a theater in he his a 1 0 1 0 1 0 1 0 1 and with that the interview moves over the image and you just might apply everything and with their with into content and put a beautiful right and then you do it again and you just move over the image with a fetus and you get a new each can watch the different you used the and of course the many
features you can ever identities future on X to computer sharpening I mean all that is in a life 4 against and you can use that and then at and and you also have
these nonlinear the rectifying units here which had to basically to learn and then
you're these pulling step where you say I I I now look at the image from a more coarse-grained perspective so in in that part of the image only take a maximum and also in the other parts so I shrink my my size down and if you
don't put all that together and you can say that the 1st part of the skin and if the feature extraction so in here and that the commands which can tests it has to use in order to extract features from image and then there are the 2nd part of the MIT's justification right so you forgot would fit to this and then at the end of right so we use exactly that friction and fall problem where we have patches and we have the annotation if that patch would become either the world class or the other right Jean and then we put that in and then again you have this feature extraction part here with the convolution McLean years the optimal features picks up I learned that you need to extract features that allow you to testify 2 pitches according to together training data set and for each of these pictures then you get a patch underscore what which tells you how much The grammatical believe that this goes to the 1 or the other now we also X displacement and this just how much did the 2nd moved and in that in that in the time between the last image and the image and mouth so the speed of the separating and more these patch features now we use and since we have tracking and we have images
and 5 minutes you see that we have a lot of these patches for 1 cell and we can just average over that image patches to get essays hike leveraged its core and by the patch is very noisy you see this patch score for us for for 1 2nd It's were to robots so this allows us to put 1 score to assess so in that case you this if you get a score of 0 comma decimal 7 and that means that our neighbor believes that this sense is a GM said and and not on the case I this energy
is the 1 protein approaches you can try to put that into a
recurrent your network and user temper information we try that as well but it didn't give us much of an advanced and now that the will be get from it this
is on this is all our data these are the experiments this is the same trees issued before in the WAN color but now also the other color so just 1 branch that of the other in there and
we used 2 experiments for training and test on the side the issue is the weather quite diverse but I was I mean they have taken different years from different pitch distance and the left so and there are variations in the lightning and the whole this as a treated and in order to do that make sure that we don't overtrain almost we do it in the wrong profession through 3 times and you get the
sum of true positives false positives ratios the area under curve as a measure for the the death of a prediction and now and we put in the
end so we train on the annotated cells those that have a mark already and then we test on the unit and says in the 3rd experiment and you see that we get a AUC is for these in it and says above 0 comma decimal 8 was the and that's quite a cake but the more interesting thing is that if we now take a so that has a total market this latent says a little bit of cation we find that after 3 durations before the market comes up we get similar a so we are able to predict the set it up to 3 iterations before this market and
since this divide these blood cells divide every 2 evolves roughly and we can go so 6 hours before the Commission markers come up to predict for the science but with the now book can we do with
its where 1 idea would be to not to
experiment with new you for the cells in culture and watch them different track them and then recover the trees but to stop that from and in between
when they haven't on the mark on it and then do maybe as things are sick experiment where you profiled expression of each of the sciences and then you with all our use of prediction to separate
the says in the 2 classes and then figure out which fact is actually I differentially regulated at this early time kinetic parameters and that by tell you this not you wanna go to that 1 which other factor for the figures are involved in this dissertation right and so Highland Heights who was initially are invited here at the meeting action boss that he would give a talk about Filani effect because that's what he's doing all the time you can imagine that things you can imagine that if you do an experiment where you measure the an expression in the bulk of the population of cells million of satisfied and you get 20 thousand data was basically which is this expression for each of the genes we had and if you do it now in 10 thousand singer says you get a matrix of 10 thousand 20 thousand in these matrices but to sort of phase with the Pomona to to make sense of it as a high-dimensional problem of the divide in in in involved in the war on data and half of another people are developing methods to be able to make sense of this large datasets clustering the cells finding trajectories and the processes and submitted what what's he doing right
as though the last part is take adaptation of what I just showed you and the idea here
came the been working together with conditions in Munich which lead gets blood samples of patients suffering from Eckart Motor the key here and diagnosing and leukemia is still based on this more fruity assessment for people were trained to do so so if you don't feel well you drop back to draw some blood and then this is smeared yet out on the last slide and settle adjusts look on the class on microscope and count since in if labor to recognizing the says it might be not nice application for them to do that it is setting right
so these are the blood so the the blood smears that that assessed by cytologist and we took 100 of those plus news from a patients and then from images which you also from and we did
start the samples in scanner basically and these guys are the red blood cells here and we are more interested in the leucocytes just go through this leukocytes and count them and classify them into roughly 18 different cell types and that's exactly what we did so we have to get a bringing search right we offer fatalities please mark 100 of these cells and patient and testify intuitive losses and then the idea was to again trainer CNN with it and see all of so
after scanning a staff and after the set doing the annotation which has not been easy because they're busy with which he of course we ended up with 18 tell some of these things leukocytes And
then we won that section sorry this scheme of the state in different classes so you know these into different as we ask the the atoms to put the
says and you can put them in this kind of
hierarchies so here are the the good so so to say hear the altogether mature says the the image was says some of them which was as I allow to shorten your blood but he's blasts here they shouldn't show up here but if if you recognize blasters in your blood and that's a clear sign of Eckart like the X-Men that appear and now the task was to identify these blasts from themselves right so this will be asked altered by to do and if we train the defi
weaknesses and then figure out how we can do that you see or sensitive use for the 250 is pretty high united that point to the craft it is the number of doesn't do much in getting better feeling of what that means and we ask the 2nd set what it used to just to exactly the same what our computed so we show that such urges the stages of a test the creator god and 2 against of occasion data different office and you see that the you with it which is performed roughly as good as our and our approach so good which targets good occasion that this can be used infer that cations in and 1 last thing you should do
when applying the sinister should show whether you learn with information or just cracked and 1 of the ways to do that is and using sailors e-mail which tell you which of the pixel you I have an the image I actually important for the front the cation and we did that here and you see the primal sites on the computer and considers those pictures inside the cell uses them as important similar to what the support with the sudden just and I I think it's important same
for that office OK I showed you that in school shooting can distill numbers of these images were and how I show that you won an existing consistent with the model
and more venomous substances is pretty different stories well I skip that I I wanna
think the people involved at all but you did the
work and collaborators my group at the
moment funding body and thank you very much for your
attention and
questions the limitations love and
Loading...
Feedback

Timings

  830 ms - page object

Version

AV-Portal 3.9.1 (0da88e96ae8dbbf323d1005dc12c7aa41dfc5a31)