 So, welcome all to this virtual SIP Computational Biology Seminar Series. Today we have the pleasure to have Bastien Chopin from the University of Geneva, and in particular from the Scientific and Parallel Computing Group. Bastien got a PhD in theoretical physics from the University of Geneva in 1988, and then he spent two years at MIT in the United States and a year at the Supercomputing Center in Jürich in Germany as a postdoc. He is now a full professor at the Computer Science Department at the University of Geneva and is also a group leader at the SIV Swiss Institute of Bioinformatics. So briefly, the Scientific and Parallel Computing Group develops new algorithms and methods to better understand and or predict various complex phenomena in biology. The group focuses on multi-scale modeling and computing, eye performance computing, cellular automata, lattice-botsman methods, and multi-agent systems as well as optimizing technique and machine learning. Today Bastien will share with us some of his numerical modeling and simulation in order to better understand thrombosis formation in cerebral aneurysms. Thank you, Bastien, for accepting this invitation and the floor is yours. Okay, thank you, Diana, for the invitation. So I'm very happy to be here, and thank you for attending this presentation. So today I'd like to talk about thrombosis in cerebral aneurysms. So for me, it's a field which started more than 10 years ago with a colleague, a neurologist at the hospital of Geneva, who came to me and said, oh, by the way, could you simulate the blood in a cerebral aneurysm? It's very important to understand the efficiency of new devices. And I say, okay, no idea, let's try. So we did some work. And then we started to say, okay, but we can simulate something, but do we really understand what's going on? And do we understand what's the effect of these devices? And then it started a long research, which actually ended up through a European project where hopefully we could clarify something which is very critical in medicine, but also with an important question in biology. So I'd like to start to discuss a bit more about my modeling and simulation techniques, because that's the core of what we do. And after that, of course, I will show how we apply this to the problem of understanding what happened in cerebral aneurysm. And as I mentioned, this work has been done by a European consortium under the acronym Trumbus. For those who are interested, you can find more information there, the project has ended now two years ago or one and a half year ago. So you might be all convinced that for me, it's been clear for many years that numerical modeling and simulation, it's become more and more critical to address challenging problem in science, but also very much in life science, where maybe these techniques are less used than maybe physics, chemistry. But also something which is important to tell is that the standard mathematical approach, which is to use a differential equation or some equation that you have to solve may not be so easy to apply in many life science problems. So therefore, there are another way to proceed, which is try to copy nature at a lower level, like what people call particle-based model. We'll see some example of cell-based model or rule-based model where you're not solving equations, but basically you're trying to reproduce the physical processes directly in an abstract way in your computer. And that's actually a very powerful way to address many, many problems. And it works well because if you're interested in a global question of your system, the detail of the microscopic interaction may be not critical provided that you are correct with some essential ingredient. So the idea is that we want to create in the computer, say, a in-silico universe in which you can reproduce your biophysical processes. And of course, the art of this thing is to capture the essential feature of the system you want to model. So just to illustrate this with this slide, on the upper line, you see a differential equation, which is what most people use when they want to solve a fluid equation. So they want to know how a fluid flows. So u is the speed of the fluid, p is the pressure, this is the viscose term, and this is a very difficult partial differential equation. What I'd like to stress here is that the steps you have to go through from the phenomena to the solution is basically these four steps. So from the phenomena, you have to build this PDE, this mathematical formulation. Then unfortunately, this equation is most of the time you cannot solve it analytically. So you have to solve it numerically. And for this, you also have to apply smart algorithms for discretizing it. And finally, you have the computer solution. You should also understand that the steps from phenomena to PDEs took several hundreds of years before people at the mathematical level to go to this type of construction. So of course, this is done now and you can use it and we should use it. But this is by no way trivial, this bottom. Now we have actually another tool which is powerful computers. And you can rethink the way to model a natural phenomena. So this is a fluid basically, okay? But the fluid, we know it's particles. It's like fluid particles that move in the space, collide, interact, and so on. So why don't we try to take this picture directly in the computer? So we'll do an abstraction of the fluid movement on a simplified space. Here, it's a 2D lattice, but of course, you can do that in 3D. But you see these little arrows here, they represent some fluid element that move either to the left, the right, or diagonally. And sometimes they can collide at the same point, meaning that they basically interact. That's what we call the collision, or it's a physical interaction. And the result is that all these fluid elements, they are redistributed into space with different velocity, like a bit beyond ball that collide. And then they move to other point in space where they will certainly meet other particles and continue to interact. So that's a very, actually, correct physical picture of what a fluid is. What is critical in this process of collision is that the amount of mass you have here is the same here. So you don't lose mass. And you have the same momentum, okay, also. And that's the important physical feature that you should keep in your simplified dynamical process. It's well known that actually this equation, Navier-Stokes equation, is just a mathematical expression of momentum conservation. Okay, so actually there's not much more in one or the other. The thing is that here you go directly from the phenomena to the computer model and the solution. So what is catching this little diagram is what people call the lattice Boltzmann method. It's, of course, a bit more complicated than that. But essentially, if you know how to explain or express this collision term, if you know how to generalize in 3D, and if you know how to put boundary condition, you have a fluid model. And that's what we have been developing for several years at the University of Geneva, in particular, Dr. Jonas Latt, who is the father of all this development, is this parallel software, which is parallel lattice Boltzmann software. It's an open source software where you have extremely powerful simulation possibility for different type of system. So I'd like to illustrate a few. So sorry if the movie will cause some problem because we don't have them very smoothly, but yeah. So this is a simulation done with this lattice Boltzmann method, a bit more than just my little diagram I must confess. But this simulates actually, very precisely, the way you drain the dam in Verbois in Geneva. Another example of this method is illustrated on this figure, which fully will start. And you can see you can generate extremely complex flows, even though it's a very simple low-level dynamical process. Okay, maybe you may have fun also to see how a washing machine works. So you can compare, you know, the US type of washing machine where it's vertical to the one in Europe horizontal and see which one is cleaning better. But I mean here we are more in a bio-oriented seminar. So I'd like also to show you this example, which is another example done with the lattice Boltzmann method through the actually the company that John Aslat created, Flowkit as a consultant for Palabos. And this is an example where you have a broken vertebra and the way to repair broken vertebra is to inject some cement into it. But of course you want know the exact amount of cement you want to inject. Otherwise, I can overflow and block the vertebra. So here is an example where you can get a very complicated geometry by medical imaging and you can use Palabos to simulate the injection of the cement and to see exactly how many cement the doctor should inject. And this is of course an important question. As you may guess, this simulation I just showed, they are not, I mean like easy to do and they might not fit on this laptop. They may require a lot of computing power and in particular parallel computing and high-performance computing mean like from 100 to 1000 of static processor actually collaborating to the computation. So where could we find this type of resources? I guess in many institutions you have them. For an engineer we have a machine with 2000 cores that can be connected and used for the same application. But you probably know that at the manic area we have Cadmus, which is a center for advanced modeling science where we have a machine here, a powerful machine with 16,000 nodes, which is actually accessible for researchers from Yves Saint-Yves, Yves Lausanne and EPFL equally. And at the higher scale you have of course the Swiss Center for Scientific Computing in Lugano, which now owns the fastest machine in Europe. So we have access to powerful resources here. Just to make a little bit more clear what it means to parallelize and since I'm going to talk about aneurysm, I'm going to use this example. So here you have a geometry of artery with this aneurysm, which is this deformation and you want to compute the blood flow inside that and you want to use a parallel computer because you have a very, you need a very thin resolution and do that in a pulsatile regime. So basically you cover this geometry with little cubes like that and for each cube you solve the three equations that I showed you with the lattice Boltzmann. And of course between two cubes you have interaction because one particle can go from one cube to the next. So all these cubes basically, they are like computing power and interacting continuously every millisecond or a second with the neighboring site. So it's of course mean that you need to reprogram code at the proper data structure. But in the picture you have here you can maybe see that you actually gain a lot of speed. For instance you can compare a situation where you have 128 processors and you can go up to 16,000 processors and basically you gain exactly the factor of the number of processors you have. So it means that your application or your way of parallelizing scales very well and you actually use well all the processors you put in this computation which is not granted depending on the application. Okay maybe a few more examples of applications we have been working on. For instance with my colleague in the hospital of Charleroi in the University of Brussels who are interested to understand if we can find the adhesion rate and adhesion aggregation rate of platelets on the blood vessel for instance. So there are experiments that can be done and basically you see here the result of the experiment. So each of this little cluster is a number of platelets aggregated so this image is one millimeter, one bilimilimeter. And then a question so if I want to extract from this picture the adhesion rate or the aggregation rate how can I do that? I mean it's not written in this picture so you have to think of a model which would produce all the process and it turns out actually that if you try to do that you realize that the current knowledge about the way platelets move and adhere and deposit is not enough to explain this image. So basically it forces you to explore more and finally where we did that with new experiment, new clinical new in vitro experiment and we end up with a numerical model which I have no time to explain here but you can see a very good agreement now between this experiment and simulation. Another example of our simulation we're working which is what we can call a cell-based tissue model. It's within the epiphysics project in the SystemX initiative so the idea is to represent all the cells as a computer entities and it interacts with the neighboring cells according to some elasticity property. So we are interested in several things so here you have an example of simply such a tissue which can grow so cells can divide and they are actually subject maybe to some physical constraint like they grow into a constrained environment and so you can simulate these type of things. Here you have another example where you try to understand how the dermal papilla of a hair on a spiny mouse developed in competition with the other cell around them and it's a competition of two types of cells which can be addressed and it works rather well. Okay so let's back to the main part of my talk which is what is a several aneurysm. I already show you a picture but this is a real medical image where you see this deformation here that it's not so frequent and even if you have one it's not clear that it will rupture. The problem is that if it ruptures it's extremely serious and most of the patients who are ruptured they have a permanent damage so then the doctor really would like to know if it will break or how to treat them so what is still unclear now is we don't know exactly why it appears and we don't know why it ruptures. Okay so there are previous projects that try to solve this but I don't think it's solved yet so we decided to address a different question is how can it be treated so suppose the doctor decided to treat it what's the best way can we give some recommendation so the key point is that what drives the evolution of the aneurysm is the interaction between the flow of the blood and the vessel. Okay so there's bio-physical constraint so what people have understood from a long time ago the clinician is that they have to reduce the flow inside the aneurysm and first they were using clips more recently coils which is much less invasive because clip you have to open the skull and that's really a very major operation coils are less invasive so basically you insert inside the aneurysm some coil of metal which slow down the flow and more recently flow diverter and if you do it well you see on this angiogram that after the treatment there's no more motion of blood in the aneurysm and this opens the door to remodeling because there's a trombus formation in there so we were specifically interested in the trombus project to understand how flow diverter works so flow diverter you can see them here in more detail it's a tubular mesh okay made of base in wires that are you know braided in a in a smart way so that's what the company do and may not tell you exactly how they do because they think that part of the secret is to have a proper braiding of these wires and basically the rule is that the the you try to rebuild the main vessel with this tube but of course you cannot make something which is totally impermeable because then you will cut all the exchange of blood with this cavity and then you make even well so you have to just reduce the flow but still keep the exchange of oxygen and and all the good molecule that will help the remodeling the thing is that this treatment has been quite successful but it still fails to work in 20% of the patient and clinicians don't really know why so this definitely something to understand and and actually it turns out that nobody really understand the detail of what happens there and what makes this part of the aneurysm clot and then remodel so the goal of trombus was to combine biological knowledge and new individual experiment with clinical observation on on the court of patient and medical modeling to see if we can get a better understanding of the process which again was not empirically to work but without a clear understanding of how so this is an example of an aneurysm from a patient which has been segmented by all kind of techniques of image segmentation and that's the colleague from you who did that and here you see that all this part is the part that clot it in the aneurysm and this part here is called the lumens so that's why the blood is still flowing so this is actually a very good for the patient to go from this to this and the idea is to produce this by inserting a flow diverter in this to help nature to do that so in terms of numerical modeling you have to realize that it's a challenge which i try to illustrate on this little diagram so this diagram show you the spatial scale at which the phenomenon take place and the time scale at which they take place and the main processes that are involved so you have the blood flow in the artery which basically range from millimeters to centimeters at the size of an artery okay and it's basically a time scale of one second which is most the heartbeat okay so it's a rather fast process on the other hand the clots which is on the same spatial scale because it occupy the vessel it develops between i would say hours to days even weeks and all this is actually triggered by the molecules in the blood like platelet and blood cells plus other molecule that i will mention which are more the order of the microns so if you'd like to put all this in a numerical simulation it's a extremely big challenge and so far we are not able to do that so we have to find tricks to try to address this kind of problem and one trick is maybe to say okay we try just to concentrate we forget about the uh a detailed simulation of red blood cells and platelet and all this we just consider that at a microscopic level and basically it means that you put all the process here at this stage here and that give you something which is microscopic the other approach to say no i want to see more in detail the thing and that's a simulation where explicitly you model all your red blood cell platelet and that's a bit more difficult but it's not a bad idea because if you look at simulation which here are 2d you can see that the this the spatial distribution of platelet and red blood cell is definitely not homogeneous in space so it means that you expect that a simple flow simulation might not capture everything and you see that there is a zone where you have a bigger accumulation of particles and other or there are less so it means that very likely we have to take care of that but so far we cannot do do this at a scale of a real patient for timescale of a real patient so then we came back with some more macroscopic simulation which is already a challenge because you should realize that this flow diameter if you want to take the real size uh this truss this little wires they have typically um a size of let's say 20 microns okay so you have to resolve all this little wires and here this is typically of the order of centimeters and you have to do that at rather high flow condition and you see it can take up to 10 days on 128 120 processor for only two uh heartbeat so and this is illustrated in this movie let's hopefully show you so you you see here uh the detail of one heartbeat and actually there will be two heartbeat and here you see uh particles which are just tracing the the flow you see that clearly this flow diameter helped producing the flow here not completely but clearly it's it's less than what you have in the main artery so is it sufficient to produce uh plot or not that uh question we have of course to uh investigate but remember that it already takes 10 days on hand the processor just to do two heartbeat so it's uh and and it's not even resolving the the red blood cells or whatever okay so here you can compare for a patient a given patient what's the shear rate uh because I will come back to this in in in a moment but uh we believe that's the the shear rate at the wall is the key for producing the trombus so you can see how the shear rate changes with and without the placement of a flow diverter so the red and the blue correspond to our presence and absence of a flow diverter so in most of this point here you see that placing a flow diverter reduced considerably the the shear rate except in a very particle point which I think is this p5 where there's some pathological um geometrical feature which makes actually the the flow diverters is locally increasing the the wall shear strength but but globally it's decreasing which is exactly what you would like to happen uh so those are of course uh patients that have been collected during the trombus project by uh an already a radiologue from uh Montpellier okay so if you do the um analysis of many patients you observe something which is very interesting and goes into this validation of this hypothesis that here you see that you have aneurysm that spontaneously trombone trombones and those that did not uh sorry um those spontaneously trombones and those they do not okay and here you have the value of the shear stress at the wall and you see that you have a very nice line separating these two experiments so it's really uh support the hypothesis that if the shear rate is low enough so below about 30 second minus one then you can promote these trombones if it's higher it's not going to happen so basically it means that for a doctor he should place a stand which makes sure that we are below this value okay so that's a result of this trombus project which already i think is an important result of course you can play the game saying now that i am start understanding a bit what's the criteria i can play with the optimization of the stand design that's something that i've been doing with colleague in japan where we are using optimization techniques to change the design of a stand to find the most efficient one with a given porosity so it's not always easy to build of course and that's of course a constraint but here it's more to show to demonstrate the possibility of doing it so now we would like really to have a full model of trombus formation which of course is based on this observation that at flow shear rate you start something and it will be an extension of a very simple model we started in 2005 with this team in Geneva where we had the 2d model which was already explaining that basically the trombus can just form partially in the aneurysm or completely fill it or sometime which is very bad for the patient start invading the parent vessel and then that's of course me block the artery so that was really a success that we couldn't explain all this observation by a very simple model so within trombus we had a biologist who could do experiment and we did extensive study and basically we decided that what looks the most likely to happen is the following set of process and maybe it looks so obvious for you but it's not written in literature I mean there's a coagulation cascade which are well known but they are much too complex to be implemented so you have really to understand what are dominant the idea and that's validated by experimental observation that endothelial cells when subject to low shear rate they produce some more tissue factor and this tissue factor produce trombine okay so basically a non-material wall where the shear rate is too low starts producing pro-coagulant material then you have fibrinogen and anti-trombine which are always present in your blood they are always created and if trombine and fibrinogen meet okay they react together they form a fibrin and that's the start of the clots okay so if you want to make the clots you you should make sure that you're somewhere in your system trombine has been produced and when this trombine interact with the fibrinogen brought by the blood it creates the clots okay but on the other hand there's also anti-trombine which is brought by the blood which try to kill this trombine so there's a competition here to try to reduce that okay and then there's a few other process that I maybe don't want to go too much into the detail now but you can really translate all this into a computer model in terms of kind of advection reaction diffusion processes so where all these letters they correspond to some particle that can interact the same place so if you do that you start having a model where you can produce the trombus and here you have an example of what may happen with parameters that may not be realistic but at least can show what may happen so here you have the blood flowing and all this black region is the fibrin that keep adding on the on the cavity of the aneurysm which is exactly what you want it's also interesting in this example to realize that if you don't take a pulsati flow as some people doing model do because of course it's much simpler to implement you would find something extremely different okay that's exact same model but with different flow condition and you see that the clot is forming where it shouldn't form in the parent artery okay so really working with correct flow condition is essential if you want to do that okay and now of course the question is can we predict in a patient what happens so we had within the trombus project some images of patient with a giant aneurysm like this one that at some point in their evolution spontaneously clot and the question is whether we can reproduce the pattern or the volume that was clotted so here I must stress that it's also a challenge in term of simulation I already mentioned this multi-scale aspect the clot in in a real patient may take weeks to months okay it's not completely clear but in the simulation you cannot wait that long I mean you see how long it is already to simulate one heartbeat so here we try to make acceleration to have the full clot happening on 2000 our heart cycle of course so it means that we actually increase the reaction rates of this reaction I showed you before so that's the way we can speed up the process but if you do that actually you observe that if you compare the lumen of the patient so this is the lumen which is left and that's where the clot took place and the corresponding simulation see a very good agreement okay so it means that we can really reproduce our observation with this model and here you have a second case where you can see the aneurysm of the patient also the clot in right which has been spontaneously produced at some stage in the evolution of the aneurysm of the patient and the blue which is what the model produces of course it's not exactly the same because there are a lot of things that we don't know it's of course several parameters are patient-specific the blood composition or this so it's not taken into account in our system so we really have a model which give a generic behavior but you see that it's already quite good and promising so in conclusion I think this dimical simulation that and modeling have illustrated which implements basic law of physics and basic law of biology so again we start to we try to start with the fundamental processes and we put all the elements into a given space over time and then we can produce virtual experiments you know in silico experiment of blood clotting and that gives you that explains scenarios or you can test scenarios and actually we were the first to propose a trombosis model it's a random that match clinical observation and in actually in addition to this we could really understand that all the the phenomenas are actually due to this spatial temporal interaction between the particles so it means that if you want to produce clout you have to have at the same times at the same position trombine and fibrin but at some point trombine can be exhausted because it reacts it's annihilated by antitrombine and so it means that you can explain rather easily a process that have been observed for a long time by a clinician but not explained is why suddenly things start why keep going for a while and why it suddenly stops so it's just spatial and temporal interaction between particles transported by blood in a geometry that may vary over time so it's of course complicated to predict analytically but you may clearly of course you can you can do that and in the end of course what you would like to provide the medical doctor with a tool that can they can use to assess the success of the treatment and of course we are still willing to go in this direction with new urban project in particular we have a new project on a center of excellence for high performance biomedical simulation that will start in October in which hopefully we can go further but there are other open project that we hope to apply for to continue this work so with this I'd like of course to thank you for your attention and also to thank some of the people who have brought some material for this presentation so people from UniG who produce the movie that you've seen my colleague from University of Lebes-de-Boselle Karim Zawi and his team the neurosurgeon from Montpellier-Omericart the team for image processing which is Gikour Bebes and his team and the team in University of Amsterdam who did this fully resolved simulation where the red cells and platelet were actually simulated okay thank you very much