 Hello everyone. First I'd like to thank Professor Joe for inviting for this exciting workshop. Actually today I prepared the two talks but the first part will be about the molecular mechanism to filter a spathier temper noise inside of the cell. I usually study some rhythm in our body so called circadian rhythm. So circadian comes from the Latin which means about a day. So 24-hour rhythm in your body is called circadian rhythm. For instance your sleep makeup cycle is one part of the circadian rhythm. So how this 24-hour rhythms are generated in our body is via transcription and translation or negative feedback loop. So there is so-called period gene. When co-op email protein, activated protein bind to this period gene, period gene is activated so that it trigger the expression of the period mRNA and then they are translated to the protein at the cytoplasm. So this happened about for 12 hours so that around the 12 hour you can see that period level is keep increasing, increasing, increasing, increasing. And then after 12 hours they enter to the nucleus where they inhibit the clock and email so that transcription is now stopped. So that increasing of this period is now stopped and then now they are degrading, degrading, degrading for 12 hours. And when it becomes now 24-hour all the period protein is now disappear in the cell so that that inhibition is also now released so that now we can reproduce the period protein for 12 hours and then decay for 12 hours. That happens every day in your brain basically. That's the key mechanism for generating 24-hour rhythms in our body. And then this mechanism has been identified by these three American researchers and who got Nobel Prize in physiology or medicine for this negative feedback loop of the circadian clock. So this is a really well characterized mechanism in biology but somehow when I read this thing I felt that this is too ideal because to generate this 24-hour rhythm this period protein need to enter the nucleus at the same time every day, right? Every after 12 hour they enter to the nucleus and shut down the transcription. But cell is not empty space. It's a super complex space. But how the thousands of the molecules enter, travel this complex environment and enter to the nucleus at the same time every day. Is it possible? Actually not easy, right? So to investigate this question basically what this means here is a New York map. Let's say different people leave at starting their travel at a different position but they travel through this old track picked jammed Manhattan road and then they enter, arrive at the center of the Manhattan at the same time every day to think it's possible. It's not an easy problem, right? So to investigate this basically what we did is we tacked the GFP protein into per molecule and we tracked the movement of this per molecule inside the cell. And then what we found is, so this is a nucleus, what we found is of course their arrival time is different. Some protein of course arrive earlier than others but surprising thing is you can see this ring structure that means they do not enter to the nucleus immediately. They are keep waiting their friends. They are keep waiting waiting waiting their friends and then when all of their friends arrive they enter to the nucleus together. Okay so they make whole this ring structure and then suddenly they enter to the whole nucleus together. So I was curious where this friendship between the molecules come from? So where this collective behavior of the molecules comes from? That's my question. So that for this we use the mathematical modeling approach. So first we use that Asian based modeling approach. So basically transcription occurs inside the nucleus and then they are translated into the at the cytoplasm and then they travel and then they enter to the nucleus and cross the inhibited transcription. That's that negative. But with the known current knowledge what we found is of course when certain protein arrive earlier to the nucleus they just getting to the nucleus earlier. There was no collective behavior. That means there is something missing in that negative period loop. So and first what it is known is when this period protein enter to the nucleus they should be phosphorylated. That's already known at the multiple site and then a couple years ago we found that with the Duke NUS group this phosphorylation occurs in a cooperative manner. What I mean the corporate manner is first the phosphorylation is very very slow but as long as that slow phosphorylation occurs the follower phosphorylation is very fast. So first phosphorylation trigger the neighboring phosphorylation. So with this cooperative phosphorylation what we found is we can make the switch. So what I mean the switch is let's imagine here is a cell and then let's make a very small circle in this cell. Then we can measure the local concentration of the purporting right. So here y x-axis mean is local concentration in this small circle in the cell. Okay and then y-axis is in this small circle what is the fraction of the purporting which is phosphorylated. So what means all of them are phosphorylated. So what we find is when local concentration is low majority of them are in the unfathombrated status. You can see that unfathombrated status it's maintained maintained but the local concentration is passing to certain thresholds suddenly we found that all these unfathombrated proteins are suddenly switched to the whole phosphorylated status. So really the collective behavior was possible in the locally. So when we merge this we call this is phosphor switch because this is off and on of the phosphorylation. When we merge it to that previous agent based model now we can finally see the friendship among the molecule. So here what I draw is here orange dot is unfathombrated protein and purple dot is phosphorylated protein so that they can enter to the nucleus. You can see that at the beginning they are accumulated accumulated at the perinucleus but around that time the concentration of this per molecule is not high enough. So that's why they are in the orange status. So that's why there are very few of them can enter the nucleus but more protein is accumulated so that finally they passing to this threshold now they they are suddenly switched to this purple phosphorylated protein and then they can enter to the nucleus all together. So this is a key of that collective behavior. Then how to prevent or interfere this collective behavior is one way is introduce the obstacle. So this is let's say normal level of the crowdiness of the cell this gray dot but that's increasing so that cell is more and more crowded. As a result the accumulation of this purple protein at the perinucleus is not occur. As a result you can see that the collective nucleus entry does not occur and this leads to the very weak and unstable circadian rhythm compared to the original circadian rhythm. So this collective nucleus entry is the key to generate the strong circadian rhythm and we wanted to validate this model prediction using experiment so how to make the cell is more overcrowded. One way is using the adipocyte basically we put the more artificial fat inside of the cell we can do it and what you can see is without that there is a normal circadian rhythm but you can see that we put the more fat that inside of cell circadian rhythm becomes very unstable and super long and here is a key experiment. So this is normal mice you can see that they sleep wake up sleep wake up sleep wake up in a very regular pattern but this is when day and night but even we make the dark dark complete darkness you can see that they maintain beautiful sleep wake up cycle but when we make the obesity mice so that cell is more crowded with fat you can see that there is a normal regular sleep wake cycle happen and then what else can make a site our cell is more crowded there are various things asing an autophagymal function and Alzheimer's disease are all known to increasing the cytoplasmic congestion level and we found that in the mice of all these three type of the phenotype have this kind of sleep wake up psychopath and then this was mice and then we studied a clinical study and we found that even in the human obesity human and aging human Alzheimer patient all of them had a sleep cycle problem but problem was we didn't know why but our study suggests that all of them has an unstable sleep wake cycle because their cytoplasmic congestion level is too high so this shows a way how to treat them to reduce to make the restore the regular sleep wake cycle and then so that so that we suggest that this this new concept this cytoplasmic tract pick gem could be the cause of the unstable sleep wake cycle so this could be a new target for the treatment of the unstable sleep cycle so this is my first talk and then let me move to the my second talk so this is about the inferring the network structure from the time-serve data it's a sort of reverse engineering problem and you know cell is complex and actually identifying the interaction between the certain molecule is not easy experiment it took a couple years in the experimental lab but nowadays due to the experimenter advance measuring the concentration change of the specific molecules are way easy to measure so that the natural question is from this time-serve data can we say that these two molecule have any interaction that's a very natural question so for this various inference algorithm has been used and one of the most famous one is like a grandeur causality or convergence cross mapping thing which is based on the predictability concept but the problem is previous previous inference algorithm has some common problem because I'm studying the circadian rhythm so module gene has a 24-hour periodic expressions so when I plug into the 24-hour same period gene expression profile previous inference algorithm stated always yes always the causation he exists because they cannot distinguish the causation and synchrony between the data set so that that means we need a some new method basically we need a some new method based on the unique property of oscillatory dynamics that's what I want to present at the remaining time so let me begin with the very simple and famous example it's a fish human argument model it's a two-dimensional OD model you can see that there are four type of the interaction between the V variable and W variable so for instance W positive will regulate the V because V dot has this positive term so let's assume that we don't know this equation so that we don't know this regulation but we just know their solution by just looking at this solution can we recover this for causation type that's my question so one simple approach is okay W positive regulate the V so maybe we can expect that as W increasing maybe V also increasing that's a one-knife approach but unfortunately that does not work even here W is increasing you can see that V is decreasing so it doesn't work oh right actually W regulate not V it's V dot right so then maybe we can say that as W increasing V that might be increasing right but as W increasing V dot decreasing so that's not also true and the reason is the V dot is not only determined by the W but also determined by the V term right so there's a masking by the V so we cannot use this kind of approach so that means we have to investigate the relationship between W and V we have to remove this masking effect from the V so how to do is very simple when we let's pick up any time point then we can find another time point TV which has the same value at this V right okay then if we compare the W value at T and TV issue the following the W so what I mean is if we look at this two value here W is smaller compared to here that means this thing indicated W dot should be smaller compared to here right so that this can provide some information so based on this we define the regulation detection function which means the relationship between W value and W TV and V dot T and V dot TV so we can imagine that if WT is larger than W TV V that should be larger than V dot TV so that we can imagine that this term should be the always positive and then if we plug in this original OD we can prove that this is true so that throughout all this time cycle we can see that this regulation detection function should be positive so that we define the regulation detection score what this means what is a fraction of this positive area in this example that's a plus one okay so if we extend this idea you can see that here V to the W is negative regulation in this case this regulation detection function is always negative so that detection score is minus one and W to W is self-regulation is negative so that is always negative so it's minus one but what about the V to V it's a mixture of the plus and minus term so that the regulation detection function consists of blue and red parts so that its value is between the minus one and one so this provide a very simple criteria if there is a plus regulation this regulation detection score should be plus one and then minus regulation is minus one and then it's a mixture it's a plus and minus one between value and then we apply our method to this simple experimental data between the pre and pre and pre data population data and here is a calculated regulation detection function P to D if we bring the P to D data the regulation detection function is plus one that makes sense because this is food and this is a data so that should be plus regulation and opposite is minus one and self-regulation is between the minus one and one because current our population has both plus and minus regulation because births rate and death rate depending on current initial population so this indeed algorithm works even in real experimental data so that we want to use for the really influence of the extra network inside of the cell so basically we made a two assumption inside of cell when X protein regulate the Y protein we assume that it's monotonic what I mean is X regulate the Y either in positive or negative way but not the mixture that's one assumption and second I assume the self-regulation is negative the reason is inside of cell protein degradation is proportionate to the current population concentration so that we can assume the self is negative so by merging these two I made a three rules if X to the Y is plus one Y to the Y is negative one so self-regulation is minus then we can say that indeed X to the Y has positive regulation and X to the Y is minus one and Y to the Y is minus one we can say that it's negative regulation but except for these two all other value we can say that there is no causation so let's apply these three rule to the this simulated data set so for instance M to the P the value is 1 and 0.8 so that there is no causation M to PC plus one and minus one so that we can say that there is a positive regulation and then here no causation PC to P positive regulation PC no regulation P to M is negative regulation so we can recover original negative group structure from this time series data so let's apply to more challenging case now we have six time series data and then we calculate all this value for each pair you can see that there are five one and minus one which made this five positive arrow in this diagram and there is a one negative arrow so that we can again recover this original model structure so that based on this we made a package we call IN so inferring oscillatory network ION package and then we apply to the some real data set this is one of the famous synthetic oscillator called refrigerator so their data looks like and then they have this three repression structure and then when we apply our IN into this original data set we were able to recover this negative negative negative structure but when we use that grandeur causality or partial convergence cross mapping most recent inference method you can see that basically all of them are connected so that they fail to recover the original structure and then I tried this one as well I copy the original data set and I shift it so that now there are six time series data so that when we applied our method we were able to get this two independent repression structure but if we use the grandeur causality you can see the basically is say that everything is connected so that's the problem of the user inference method when the oscillatory time series data and then finally we also apply to this only complex system and again oscillatory time series data with our method there are only two regulations to regulation is detected but if we use other previous I would you can see that basically it predicted everything is connected okay so here is a summary what we did is basically when the time series data is given we asked the question whether this kind of ODE exists with the monotone function of F which can reproduce this oscillatory dynamic series if the yes then we say there is a causation if no then we say that there is no causation that's our model based inference algorithm then of course one can ask a question may what if we generalize this model right so that here we only consider that for the wide edge only single input right but there could be multiple input cases for to answer this question we have to use this generalized model and then fortunately my two talented students were success recently successfully extended our approach to this gender most the gender type of ODE so that when we apply this to this time series data you can see that now more complex that truck structure can be inferred and then and good news is now the new method can be applied for even for non-oscillatory time series data so any type of oscillatory any type of time series data can be applied for instance if we apply to this air pollution and cardio patient data in Hong Kong we were able to found that among the air pollution NO2 and this I don't know what it is but this two air pollutions are key for the cardiovascular disease in the Hong Kong population so let me summarize talk first at the first of we are we found that in the circadian clock search certainly some protein arrived to the perinucleus earlier but they are waiting there because they cannot be phosphorylated but they are more accumulated so that the threshold is past then suddenly most of them are phosphorylated and then they can enter to the nucleus and this is the key to generate the strong reading and then filtering the spatio-temporal noise of this arrival times even arrival time is very heterogeneous due to this by stable switch which they can enter the nucleus at the same time that's the first part and the second part we basically from the time series data we asked the question whether there exist a certain smooth model which can reproduce these data based on this question we were develop an inference method for the coding from the X to the Y and we name it as a general model based inference method and so the first circadian clock project is done by the sub to Chae and they were Kim now who is up and new assistant professor at University of Michigan and then the second causation project is done by sale Park and Sungmin Ha and as a professor Joe introduced me I launched this IVS biomedical mathematics group last year it has been an ear it's very new center here basically various background people math people biology and physics and chemistry medicine people work together to solve the biological problem and then actually we are organizing the online colloquium series it's a public event so if you have interest in this system biology or computational biology stuff so you're more than welcome and and if you miss to your the talk we also have a YouTube website you can see the our previous talks so yeah and finally we are hiring the senior researcher and post doctor so if you have interest in this kind of stuff please let me know yeah thank you any question yes and great talk I just have a couple of questions yeah if I took randomly selected intervals mm-hmm out of the time series mm-hmm right and I just treated them as independent variables except I demand the coefficients in the functions mm-hmm are the same between the different time intervals right would I get the same result or not because then I'm not I'm not claiming that I need to use the fact that there is a point at which that right yeah but at the same time kill our new approach okay so in that way we don't need a oscillatory dynamics so I mean that's really what you want right regardless of the time interval it should be the same function right yes yes exactly okay so in that way we don't need that selecting the two time points thing so that can be quite restrictive yes especially especially if it's a noisy system mm-hmm mm-hmm okay second question did you check the how did you quantify the the crowding thing because adipocytes when they get big mm-hmm unfortunately I have a lot of experience studying adipocytes size and and when mice are obese or rats or whatever you want basically adipocytes are very eight typical cells oh yeah yeah oh so I am just curious how you quantified crowding because in an adipocyte that's full basically all the cytoplasm is a tin sheath around a big blob of fat right I see yeah so for the adipocyte experiment at the in vitro experiment because I do was we make the adipocyte is keep increasing so that we are we are pretty sure that as time goes the cell becomes more crowded that was made sure a flea is sure but for the mice case it's just obesity mice so that obesity cell crowd is not just by the adipocyte there could be multiple thing happen but unfortunately there is a no way to actually measure the cytoplasmic congestion level at this point with the current techniques so every reviewer actually ask us can you map can't quantify this this but because fortunately there is no way to measure it at this point yeah no that just imagination yeah okay because the obesity in the actual mice is going to bring along with it all sorts of other complications like local inflammation for instance and so on so I think it would be very hard to distinguish biologically what exactly is going on yes so I have a live question in the end you are studying the dynamical systems and so you are periodistic of your time trajectory I was wondering how long it is and then I was wondering if you see some evidence of chaos in this in your time trajectory have you have you actually tried to look at the one car section something like this the first one or second one in general I mean you in first one second one it doesn't matter actually anyhow studying the dynamical system I'm first thing I want to see is the if you see the some period but it's not perfect yes yes there are some shifts maybe due to the noise but they get the punker section from there so kind of for the simulations today we use the first of perfect oscillatory series but and then we added some noise so that there is some kind of peak-to-peak is changing and amplitudes change but actually the pattern inference pattern doesn't change much even what that happened but for the real data set unfortunately we cannot test that kind of thing because if you look at the real data set it's usually just two or three cycle so there is no such long data series so that we were not able to test that kind of thing but just what we know is not every cycle not exactly same so there is a slight changes yes yes and you also saw that the method can be extended to the K with multiple input but still like how to say just input is a superposition of the total input is a superposition of pair why input can you come in on draw like way on your intuition like if you have high order interaction okay yes yes very good questions so let me begin with the one example for instance let's say where is multiple input cases oh yeah for instance this one this one has a two input right let's begin one input case our first this red a has a one input right then if this one first this one input is true then what happened is if I add additional things we found that our algorithm says always there is a causation to reason is basically we are testing computer yes actually this is a single input cases but let me ask this question you IDT equal to and this is actually yes case and then other variable if really x causing the y if this multi-dimension there are od exist then of course the answer should be yes because all these other variable coefficient could be just to zero right there are these things in between the direct input so yeah so direct input here in direct one because you have a network system so you have a multiple pathway going from one node to the other node so if the method it just like how the method work when you have the when you need to take into account multiple pathway this is a question so regarding that our algorithm distinguished the indirect and direct input it can it in our algorithm indirect causation is not considered as a causation only this algorithm detected direct causation if you now to have a x1 to xn all of them have some influence on y but what can happen is in you you have some kind of high order relationship like x1 multiplied x2 multiplied x3 and the combination of that is also okay because x1 x2 x3 is also part of this model this f can be any general function so it doesn't depend on any specific form of the regulation function as long as it's continuous function okay so that's why we call it a general or the general model based approach rather than specific model based approach so it doesn't depend on it could be heal function or mckelsmann to function polynomial function sign function whatever is okay yeah yeah any other question I want to ask you second talk so for example for oscillatory system I think we need some continuous data and I guess the required time interval for the data would be one period or several periods but however for the non-oscillator system we could not determine how much time required yeah so what is the solution for this okay so for the first part of course we are experimenting the data they are not continuous right so what we did is for this discrete time point we just interpolated and then we treat as just continuous data and then it worked so that's for the first question and for the second question for the oscillation we know that where we finish but for the non-oscillator system of course if we have more more data it's always better so that our second algorithm what is newly added is with the current data can we tell or not there is what first algorithm say that from this data we cannot tell anything so so that we need more data there is a we add that part so that how much data we need to tell the causation or not so for this case is as long as one oscillatory time series data is given we can clearly say yes or no but for the non-oscillator time series data in our new project a new algorithm we add an algorithm say that current data is enough or not first and then we if it's enough then we say yes or no so that algorithm output consists of three yes no not enough data set yeah okay thank you yeah other questions okay fine yeah what happens if there is a time delay yes very good question so that's the big issue currently we assume that the observed time variable at the regulation is immediate but for certain level of the time that it's okay but the problem is if the time delay is super long then our equation is ODE equation right so that we cannot use it so that actually our follow-up project is now we extend this to some DDE equation and then we are trying to figure out this issue yeah thanks for the question yeah so you have another question i like you have a lot of questions okay shall we start here let's thank professor kim