 Good morning everybody welcome to the third day now of our summer school So as we said like in the first day, we were looking at some Metaphysiology to try to understand the basics Yesterday we were looking at imaging image analysis data analysis that you need and today We will start with really like the core modeling. Let's say it that way. So talking about the models And it's my real pleasure to welcome Blanca Rodriguez. She's professor of computational medicine in the University of Oxford and For years. She's been working on models, especially of electrophysiology Cardiac electrophysiology at the cellular level, but again as we try to emphasize in the summer school It's like always this going back from data to models to Experimental models and try to iterate in this like never look at things in a very isolated way And this is exactly what Blanca is doing. So it's a pleasure to have you here. Please. Thank you Let me just check out. This is working well. I'm not too loud. Yeah right so I Will be talking today About research we do in Oxford and I will try to to give an overview of what? computational models have Delivered in cardiovascular science with an emphasis in cardiac electrophysiology, which is my field The the main emphasis is as Bart was saying in Explaining how computer models need necessarily to be linked to a research question But also to the experimental data that can be obtained in in any and a particular field so For me, these are the main challenges in biomedicine so the main challenges the main challenges in explaining phenotypes and especially in disease and I think we need to understand computer modeling in light of these challenges Because when we are asked what has computer modeling delivered We are always assessed in relation with other areas of biomedicine experimental in clinical so The for me the main challenge in biomedicine is in explaining phenotypes, especially in disease The links between genotype to phenotype and we were just having a conversation about that Understanding the interplay between structure and function very often we have different modalities assessing structure and function And it is very hard to link them both Understanding population heterogeneity and I think increasingly there is science pointing towards the huge importance of environmental factors in determining phenotypes so genotypes is easy to measure Understanding experimental factors is really really hard, and I think modeling and simulation has a key role to play in understanding all these challenges So another challenge in biomedical Biomedicine is the huge variety of biomedical data in cardiac electrophysiology We have in vivo and non-invasive measurements mostly from imaging, but also Functional data like the electrocardiogram. We also have in vivo invasive recordings from catheters and biopsies and The difference between non-invasive and invasive is that the actual method of recording it is having an effect on the Experimental data you are obtaining and that needs to be taken into account in the interpretation We have ex vivo recordings and in vitro recordings The most important thing for me in light with the challenges I highlighted is that these data provide one snapshot on the time of the recording They don't Characterize the phenotype preset because that changes over time very often especially for the heart But they provide one snapshot. They are also multi-scale So often they give you information at one particular scale and either on structure on for our function So it's a huge challenge to actually interpret these data sets And there are a variety of techniques that are used in biomedicine when we are assessed Our grant applications are assessed for funding. Usually we are evaluated in this context We are evaluated not in the value of modeling and simulation to say But they we are evaluated in is there any other way of looking at this problem that would be more fruitful and This is something that can only be seen when you are in panel in panels for funding so there are different ways of looking at the data and The very accepted one is statistics so medical doctors are trained in statistics And that's one way of doing it image analysis and signal analysis are also well accepted and Modeling and simulation is starting to be more accepted But we are still a minority and we need to demonstrate what is the value of modeling and simulation Compared to other methods of analysis increasingly we are seeing examples of use of machine learning and Also crowdsourcing so these are becoming really fashionable In science we still need to see what is to come and what can be shown they they are producing There is also because because deep learning is also very fashionable More people are are trying to apply this deep learning methodologies to biomedical data And this is another area of increasing interest Now what I would what I will do because we are in the VPH summer school is to focus on modeling and Simulation, but I think it is really really important to emphasize that when we are evaluated We are evaluated with respect to all these other imaging and signal processing techniques So if there is a better way to look at the problem don't use modeling and simulation Because you people will ask you why why using modeling and simulation for that even if it's your background So I looked in the PubMed I did a PubMed search and I looked at how many papers have been published with Computational model mathematical model or computer simulations and in silico in the last years and you can see here an increase in the number of publications I wanted to have a feeling for how many where specialized publications or publications in specialized journals or publications in very very high impact journals where we are Showing very new research findings that are likely to have a huge impact So of all of this a small amount are in this sort of publications I'm not saying these are the most valuable ones But certainly when we are assessed with respect to other fields We need to prove that we are at the cutting edge of research And I wanted to see how many were published in journals that are considered to be cutting edge So there is an effort to be to be made to move modeling and simulation at the forefront of research in biomedical science so In in cardiac electrophysiology Everything starts in 1960 with an is novel actually publishing in nature the the first model of the cardiac action potentially was a very simple model With sodium and potassium currents, and he was able to simulate the cardiac action potential This is this is an example of the computed action and pace Make your potentials and the the ionic currents. This model was absolutely wrong He was completely wrong in that he only assumed two currents sodium and potassium to be active in the cardiac cells We now know a lot more about this kind of action potential and in fact in 1960s we know that the action potential in the cardiac cells is the response of the cell to an electrical stimulus and it is Due to the opening and closing of a variety of iron channels in the membrane so this now we know that the the cardiac action potential in the in the in the human cells is Due to a large number of ionic currents that are Produced by channels opening and closing randomly Due to an electrical stimulus that is delivered to the cell The thing is even the even though the Dennis noble cardiac action potential model was absolutely wrong in its assumptions of only two Coverns it has been extremely useful in the discovery of these iron channels It has been used over the years to show how some of these channels are needed to reproduce the behavior of human cells And this is I think is the most important role of modeling and simulation in a constant iteration between Experiments in simulations to aid and accelerate the discovery of new mechanisms so At present this is the picture of how cardiac models are produced in cardiac electrophysiology We have we start usually with ionic current measurements usually with voltage clamp And these are obtained for each of the currents that I just show in this slide before These these ionic current measurements are far from being non invasive so the first thing we do with the cells is to isolate them and This creates a lot of pressure on the cells and some of the iron channels in the cells are destroyed so The single cell isolated from the tissue is completely different to a single cell in in the impact tissue and The first thing is to know that these ionic current measurements are just an Approximation of the kinetics of these ionic current and we just don't know how they operate in tissue The first assumption we made when we constructed computer models is that the kinetics of these currents are intact So we trust them But the actual conductance of the amplitude of these currents would be very different in In-tax tissue and that's an assumption we made because otherwise we wouldn't know how these iron channels operate So in terms of equations, this is the equation of the sodium current that is present in all the excitable cells It has a maximum conductance That is the product of how much current flows through one single channel and the number of channels the gates and a driving force that depends on electric field and concentrations graded so these kinetics are Determined by these recordings and we trust the recordings in giving us information on this part of the equation This one here depends on the number of channels and it's very much affected by the isolation procedure That allows us to record these cells. So this one is going to be available and it's unknown and cannot be measured directly So once we have an equation for each of the ionic currents We'll put it together in a system of ordinary differential equations that allow us to simulate the action potential And then we have a way to compare with other recordings that can be obtained either optically or through microelectrode recordings now Until very recently people were Proposing a single model of the action potential of the cardiac cell so we were all assuming that One Simulation that was deterministic of the action potential would equal the action potential of all the of the cells That we all have in all our human heart and this has been going on for a long time The the way we we simulate the propagation through the through the heart is by using an established model That is the bi-domain equation which is two partial differential equations coupled through this action potential model here And this allows us to simulate the electrical excitation through the ventricles or the atria and there are different ways of Approximation from the bi-domain model, which is quite time-consuming also because the problem is very stiff because of this sodium current So we can approximate Depending on what is the research question we want to answer by monodomain Icon or graph-based methods and we have compared these methods and people have used them for different different things So technically we know how to simulate the electrical excitation in the whole ventricles This actually gives us a very good link with some imaging modalities and especially Anatomical models so because we know the function from ion to whole organ We can also obtain information from in vivo Electrograms and MRI or CT scans in order to link structure and function, which is a very very important Aim of the simulation by the way, you can stop me if you want and if there are any questions so the way I've explained computer models Until now is as a representation. So we are trying to represent the ionic current in the cells We are trying to represent them the propagation of electrical excitation. We are aiming to represent the anatomy of the heart and This is one way of looking at the models that allows us to develop new models and new techniques In fact, the field has advanced very very much in terms of representation and what I'm showing here is two examples of papers that are pull that Have pushed the boundaries of what we can simulate in terms of the heart and have Focus not only on the electrical part, but also on the mechanical and the hemodynamics and This way of understanding models as representations as way to capture reality or to capture the data is one way that allows us to discover new ways new new ways to new methods and new Mathematical techniques and numerical techniques that allow us to Simulate increasingly complicated functions of the heart So This is what I would call the first type of simulation studies The ones that aim to build models and develop techniques and tools and those understand models as representations The the main aim here and it's important to understand is technology development to push the boundaries of what can be done computationally and the outcome is technology itself was is mathematical models is numerics They identify needs for new developments like adding the pockinja system or cardiac mechanics or fluid dynamics they are very much based on computer science and mathematics and They provide a toolkit that allows to simulate to to They provide the opportunity for simulation studies to answer biological questions, but they don't answer the biological questions It's a way of understanding models as representations in a domain that is computer science and mathematics what I want to emphasize is that that's not the only way of doing modeling and simulation and another type of simulation studies is the ones that Aim at answering a biological question and that's very much the context in which we work in my group Let me just drink some water So this second type of simulation study Very much starts with a scientific question that we want to investigate and that Needs to be evaluated based on experimental techniques that are available and experimental data That are available to build the models which was the first one To evaluate them against experiments and I'm going to avoid the term validation here And we could have a discussion about it if you want, but I'm going to avoid the term validation Which probably more of you are used to and I will use evaluate or even qualify which are alternative terms for validation That raise less expectations Okay, so we we build the models and evaluate them based on the scientific question We very much want to investigate and if there is much in the evaluation between the experiments then we investigate This research question and we may propose new experiments if there is lack of much for me It's always an opportunity for discoveries. What is missing in the model? that Provides this lack of much is it that we haven't built a model with some mathematical techniques or is it that's an Important component is missing in the model that we need to incorporate So this is the framework for the types of studies that we evaluate in this in this framework here The models are not only seen as representations and they they have a Representational component when we build the models and we evaluate them, but they are seen as tools to discover something So there is always these two sides of the coin in computer models in biomedicine one is they need to represent that the real value is what they help discovering and and How they are seen as tools for discovery and I would argue that the greatest discoveries have have come when models were far away from Being accurate representations of the system, but they were simple models that were used as good tools very effectively So this is an example of this simulation study type 2 that I did when I was in the US with Natalia Trianova We build a model. I didn't actually I was very lucky when I got to the lab because Jamie Eastern had Developed with some students like Felipe Aguil Computer model of the rabbit ventricles. It was the the first by-domain model of the rabbit ventricles and It allowed us to investigate the effect of electric shocks on the heart And this is this was very very important at the time Because there weren't any studies on the effect of anatomy and the effect of electric shocks which is very relevant to defibrillation studies and We were collaborating with Igor Effimov who was doing optical mapping studies So we had a way to compare the rabbit model to the experiments So with the model was built and we had to evaluate how the model was performing Compared to experiments now Interestingly, I received two two dates or two two data sets from Igor one was early experiments And they actually agreed really well with my simulation And then I got a second set of experiments and they didn't agree So the second set of rabbits had a rhythmic like crazy so you you you could generate that with me as in those hearts very easily compared to the first ones and I asked Igor what was the difference between the two and he said that the second set they were much older So it turns out that the older rabbits were very prone to arithmias compared to the young one So I thought why why is that that my model agrees with a young rabbit rather than the old one And the reason is because my ventricular model was very homogeneous very smooth It didn't have any fibrosis. It was all young So in reality what happened is that my model was a model of a rabid young rabbit heart Which didn't incorporate any fibrosis or any heterogeneities and I learned a bit more about my model so Actually, if you see a rabbit heart You you would see how how that is using experiments You would see how smooth it is compared to the human hearts that are used that you can see in the clinical In pictures, so it's actually a very important difference between the animal experiments and the and the human hearts So I I learned what type of experiments I had to use to compare my model to and I learned through the process What we did is was quite complicated and I won't go into detail into it But what we were doing is comparing the experiments with the simulations and looking at How similar the the effect of an electric shock in the heart was between experiments and simulations and actually It was quite consistent and what we were looking in This evaluation and in this comparison between experiments and simulations was looking for consistency How consistent where the the results of my simulations compared to the experiments And what I was doing through it was building credibility in my model, which is a very important a step in the evaluation of the models and in any publication that is of this type We need to have a section not on validation per se because we will never be sure that the model agrees with reality And actually that's not the aim, but we were looking for consistency and increasing the credibility of the model So in this case we went forward and we could prove that the model and the experiments agreed and actually one critical difficulty in all of this was here we had one model but we had different experiments and The data analysis was critical to show consistency if we had five models all give of five Experiments all giving different results. How do I compare with a single model that gives me a single result? And I spent ages producing this figure that showed consistency But we have very different frameworks in experiments and simulations And this is where I started to be interested in viability and heterogeneity in different hearts So in the validation evaluation qualification of the models what we're looking for is consistency and Credibility of the model simulation and experiment system So when the the comparison between experiments and simulations fail It's not the model that is wrong is the model simulation experimental system If I had to be comparing my model with the experiments with all rabbit hearts It's not that the model is wrong is that it doesn't match the experiments that I'm comparing to So what we put forward in a in a collaboration with my friend Anna Maria Carusi who's a philosopher of science is that the evaluation of the models is not an evaluation of the models per se Is the evaluation of a system that is the model simulation and experimental system that needs to be consistent in order to build credibility it may be that the numerical methods are fluid Not necessarily that the model is wrong I think that's a very superficial statement that we very often see in after presentations of models and simulations, so This evaluation for me is critical and it needs to be considering the model simulation and experimental system We then once we did this we build a model we evaluated against experiments Then we went on to investigate what were the mechanisms underlying some of the phenomena we had seen in the models So in this type of studies we've milled the model we evaluated against Experiments and then we add value what can modeling and simulation tell us that we couldn't do experimentally in this case The the experiments had a big a big Limitation they only provide information on the surface of the heart So with modeling and simulation we could provide information of what happens inside the heart And that was the added value we were obtaining with the models and the simulations So that actually is very very important when we want to publish high-impact papers We need to build a model we need to evaluate it against experiments if there is much we investigate the question And we need to in to provide with models and simulations something that the experiments can't do in In this case it was to look inside the heart in some other studies It was about ionic mechanisms, but it is usually about understanding mechanisms of something that is observed experimental This is the third type of simulation study that we would propose Which is when there is a follow-up when the predictions of the model are tested experimentally so in this case We build a simple model a 1d fiber of the cardiac cell and we wanted to investigate The adaptation of the QT interval with an increase in heart rate So Clinically it has been shown that when you go up the stairs your action potential duration shortens or your QT interval Shortens and that the way in which it shortens the dynamics are related to sudden cardiac death So people who don't adapt very well to exercise are at higher risk to to die suddenly so That has been shown clinically and we wanted to understand what were the mechanisms underlying this process Why this QT interval shortens and why people who are at risk of dying suddenly would have problems adapting With the adaptation of the QT interval. So we build a model and we compare two experiments and We saw that the action potential duration in the experiments It does indeed shorten just than the in the simulations and in the experiments and we investigated the mechanisms So through the modeling which I won't go into the detail We show that it was the sodium dynamics that led to the shortening of the action potential duration And in fact it all dependent on one ionic process the sodium potassium pump We could have published that paper in six months and a step where you was leading the research So in six months we build a model we conducted the simulation studies and we identified that the sodium potassium pump was the key player in this But what if the model is wrong? We had no way of evaluating whether the sodium potassium pump was indeed critical or it was a model art artifact So what we did is we asked our experimental collaborators to test it we came up with an experiment that That could Test whether the predictions of our model were true or not and what they did is to block the sodium potassium pump And they measured this APD adaptation So in this study we went from building a model Evaluating it with respect to what we wanted to investigate and then testing the predictions of the model with other experiments And so these took another one year and a half And as there was very very keen to wait for it to happen and the paper was much more interesting and much more grounded in fact The first question that they asked me when I went for interview for these fellowship that I'm holding at the moment Whereas can you give me an example of a study like this where you build a model you? Investigate something it leads to a prediction and that actually is tested Experimentally can you show me that in your research? You've done that because so many of the modeling studies are about I build a model I show something and I publish it some somewhere and It's never tested so this type of approach where we investigate something and then we test it Experimentally is what more biomedical scientists would see as of value Extremely challenging rather than six months into three years, but if you have the time it's worth it so When planning a simulation study, I have seen eternal mornings, right? We never have enough data that data are never enough and they are usually not very good compared to what we expect as people trained in mathematics and computer science and engineering so There is an eternal morning. I've seen over the years They're never enough data, especially if they are clinical and they involve patients. They are hard to obtain, but I assume nobody would like Clinicians to play with your heart very much So the numerics are numerical approximation So there is no exact solution and there's sources of uncertainty in there too models are wrong that I see all the Time when people come with a mathematical background models are wrong. So what? Yes, they are So there is always a trade-off between important and feasible in the world of imperfection Okay, so we are dealing with imperfection in biology We are dealing especially if we deal with clinical data, but what is it that we can do to improve things? It's a question. So when when I plan a steam simulation study, I always ask questions Like why is a simulation study needed here? What are the limitations of the experiments? We want to overcome with a modeling and simulation study What is the scientific question and why it is important if it is not an important question go and do something else? If it's only a study that you can do and it's easy Ask yourself whether it's really worth it because it it can always take time So then once you have the research question Define the model the simulation and the experimental system and their limitations Evaluate the computation and numerical feasibility and that's always a given the numerical methods need to be sound Define the experimental data needed and available for each step So if you're gonna think of of what is the best possible outcome of this modeling and simulation study? And do you have the means to test? The next step do you have the experimental collaborators that will allow you to test the predictions of your model if you don't? It's okay. They might appear in the future But I think it is worth considering whether the predictions can be tested at least and Try to anticipate what the best outcome of the study could be so if you discover something amazing is it really going to be truly amazing and Identify the type of experiments that could confirm your findings and who can do them I think experimental collaborators are key and they are very Very difficult to nurture. So it is also important to know how to handle the collaborations We were talking about Authorships in papers and so I the more and more I I deal with these things I deal with them as with the companies with Industrial partners when we actually even write agreements. What's the project about who's going to do what? Who is going to be co-authoring the paper? And it's almost like the way we work with industry is actually very good to work with Experimental and clinical collaborators to actually have very clear agreements from the beginning So once you have done all of these usually you need to refine the scientific question because you you are overly ambitious Or not very ambitious So it's an iterative process. I if you haven't come across Uri Allen, I really Recommend his papers on how to choose a scientific problem He's a genius communicator and he has tons of videos online on systems biology And he has very very good papers on systems biology in general This is a figure from one of his papers that shows how you think a Scientific project is going to go from A to B straight line and actually how it usually is and you end up in C And you don't know how you you You got there Especially that's true for PhDs, right? So I'm going to To illustrate a bit what I think cardiac modeling and simulation have delivered so far and what are the challenges ahead? But in light with the challenges of biomedicine that I saw I I showed before so I I Like this this picture of the dance by Matisse and I do think the more The more I think about modeling and simulation that the more I think it's a dance with Experiments and clinical data and that is how it needs to be understood So this is again, then his novels model and he has if you're interested in in Iterations between experiments and simulations Then his novel has a a paper in hard rhythm that was a lecture that he delivered Where he talks about how the models have helped discover new things in cardiac cellular electrophysiology and He already talked about The iterative interaction between experiment and simulation that will we will gain and that understanding of other human heart It's a really important paper. I think and one that I always give to people who say models are wrong and Yeah, so this is again What I showed before and how in that paper then is describes how the iteration between models and simulations has helped Discover new iron channels. I will refer to this. I I have also this I don't mean to to say that Dennis was pretty historic I mean is to say that these models were a Representation of reality that was useful at that time and that has been useful to discover something that looks more realistic But it's still a tool This is the picture of The different type of models that have been produced in the meantime so the you can see here the names of the first authors and the dates and This shows the animal species. So they are huge amount of models now developed for the action potential of the cardiac cell and The assumption is always the same that the promise that is unique. So in all of these models It's always been identifying the parameter values as unique Deterministic values. So again the action potential model for the rabbit is a single thing That would be the same for all the rabbits of all the cells in all the rabbits all the time in the world so The the contribution of these models has been huge And I would highlight this model because I think it's a very very important milestone This model is the model that Tomohara produced with your amputee with data from András Barra and last love you act It's the model of the human ventricular cell And it has it was developed using human data They spend a lot of time Obtaining this human data. I highlight it because it's the first time the that's the regulator The you the FDA is considering the use of a modeling of a model for regulatory purposes So if you're not aware of this The food and drug administration which is responsible for the safety of medicines in the US launched this SIPA study for pro-rhythmic for pro-rhythmia For pro-rhythmia safety assessment So what they wanted to do is to replace the thorough QT study, which is a study based on Healthy individuals people like you and I go to the clinic and they take drugs and they measure the QT interval In order to assess whether I say for not so in July 2013 The FDA announced that they wanted to replace this clinical study through by the use of a model study and The use of stem cell derived cardiomyoside the I was there So when when they announced it they announced it one day to pharmaceutical companies and the next day they had organized Meetings with experts to see whether that was possible or not So for me it was the other way around but it was literally like that on the 23rd There was the meeting announcing to Pharmaceutical companies that they wanted to change things and introduce modeling and simulation and everybody was really scared I know everybody was complaining a lot and the next day we had a meeting where we were assessing whether that was possible or not So for me, it's a milestone because what the FDA wants to do is use this model The ohala voodoo model in particular They think it's good enough and we know enough of the cardiac electrophysiology That we can use it to evaluate the safety of drugs now We didn't think all the research had been done to prove that this model had everything it takes To assess whether a drug is safe or not But we have been doing a lot of research and we are doing research with them And it does seem that the model can be a very useful tool to predict the safety of cardiac drugs So I think we have reached a milestone in cardiac modeling where things are not just about scientific questions and Discovering new things and answering new questions It may be that we are going in the direction of using cardiac models Instead of clinical studies and healthy volunteers and instead of animal studies and I think it's really really important So if we look at the citations of these models here of these papers You can see that some of them have been cited hundreds of times So what has this this scientific research led to so they are over a thousand models who have a thousand papers Who have used the lower voodoo model? So what are they about what have we discovered with them? One of the key key topics that I think was pioneered by Colleen Clancy and Yoram Rudy is Linking genotype to phenotype. There are tons of papers that have had a very good contribution to understanding How genetic mutation and especially those affecting channels chelopathies can affect Can can lead to a phenotype that can be assessed in the clinic these linking genetic defects and cellular phenotypes This paper is a 1999 paper, so it's a long time ago But since then a lot of papers have done this type of research and they are becoming increasingly important Because of all the research that is being done in genomics So one of our roles will be to integrate the knowledge that is Obtained from the genetics of the cardiac disease and any other type of disease to explain phenotypes This is Another study that I really recommend reading which I think is also another milestone in In the use of cardiac models this study made use of clinical database combining two elements Database of 633 subjects with a mutation Called LQT one Sorry, yes, so the subjects had one mutation and this database include 34 mutations and a model That was quite simple actually, so it was a 1d fiber model like you see here, and it allows to Simulate the propagation of the action potential through the fiber and calculate the transmittal dispersion of repolarization so quite simple and in fact the people here like Jeremy Rice and Matthias Royman they have the means to compute much bigger models They have the means to compute whole-heart models with electron mechanics and everything, but they chose a quite simple model For this project and one should ask so we should ask ourselves why they did that Why having the possibility of computing huge cardiac? Electromechanical model they chose a simple model that was a 1d fiber to calculate Transmittal dispersion of repolarization as a way to predict the clinical outcomes. I Haven't discussed this with them, but I would assume that if they did it there was a reason and one reason is that The clinics need simple solutions so If they didn't go for a more complicated thing It's one because this might be enough a 1d fiber model that can run in a desktop But also because a single model that can run in the desktop Could be used by doctors and the big electron mechanical models at the moment Can only be running for computers and that's complicated for doctors So this paper show that using a 1d fiber model and the information on these 34 mutations We could predict clinical outcomes Quite efficiently, so I think it was a very important paper showing multi-scale modeling linking mutations to a clinical phenotype and That made it to very a high impact journal This linking Genotype and phenotype has been shown in other fields and it's because it's a challenge in biomedical medicine We're going to see a lot more papers like this. This is on tooth So it's a on teeth and and the development origins of morphological variation So it's a letter in nature that also shows the Mathematical model linking genotype to phenotype another important contribution of Cardiac models has been in the understanding of ventricular fibrillation, so Ventricular fibrillation is the The main cause of sudden cardiac death So when people drop death just if they are athletes sometimes it happens Because they have a mutation because they have a cardiomyopathy. They didn't know about Very often it's because of heart attack is happening. So the occlusion of her coronary artery and Ventricular fibrillation happens because spiral waves develop in the heart So rather than a very synchronous contraction things start to go nuts and the spiral waves start to occur in the heart So the first time that this was shown Was in a mathematical model was in 1946 and in the in very very early on people were showing that Mathematically you could predict the occurrence of these spiral waves It's only when they discover optical dyes in the 1990s that people in Halifax lab Starts into measure these spiral waves in the heart So the mathematical models preceded the Experimental recordings by 50 years and Because they were understood they were also able to interpret them in the experiments Rick Ray is now in the in the FDA actually so he's a He's pioneering the use of modeling and simulation also for devices like defibrillation But he was an academic doing optical mapping one of the first optical mapping studies of Reentry in the heart The fibrillation is another field that has been pioneered by modeling and simulation So it's the application of very high energy electric shock to reset the heart and we published a paper Natalia Toranova Where we talk about how modeling and simulation have helped Understanding the effect of these electric shocks. It's actually interesting because there was a lot of research going on in defibrillation Some years ago, but things have stopped a bit and I wonder why people have stopped Understanding they're trying to investigate the effect of electric shocks in the heart I did a lot of research on defibrillation and one of my difficulties and why I stopped is because it's very hard to obtain data that can be used to In conjunction with the models and experiment and simulation So it was kind of easy to do research with by domain models and it's becoming increasingly easy to do simulations of by domain Models, but it is much harder to actually get data to Evaluate those predictions and it's an area where there is any huge clinical need But not very many very many people are studying it It was actually in in the 1989 when the first simulations of the effect of a point stimulation were showing how the Polarization of the tissue when when a point stimulation was applied was very complicated And what they showed in John Wick's was laugh is that the effect of the electric shock on the heart depended on structure So what the simulations were able to show with a passive model? So not even an active model with ionic currents was that the effect of the electric shock very much depended on the fiber orientation and that set the scene for a lot of studies that Included these experimental studies showing the same thing as the simulations Where they show that the polarization of the tissue and the electric shock would depend on the structure this Discovery was hugely important Because it means that people who have an abnormal structure would have an abnormal delivery of the electric shock And it is very very important to understand structural abnormalities in the heart to understand defibrillation and how a defibrillator can be Effective so again in this case it was modeling that show predictions that were matched by experiments in the same lab Actually a few years ago these people are physicists So for the physics for for people who have studied physics It's pretty much the same to do a mathematical model than the experiments using optical mapping They are used to and trained to do Experimentation and mathematical modeling so I think for them it was much easier to construct a model show predictions and then do it the experiment themselves for most of us That's high. That's quite complicated. I have a computational lab and I have I don't have the means to do the experiments So it becomes a challenge of social abilities to actually build the Collaborations to actually do these type of things where you want to test the predictions of the model in the lab yourself So it was quite impressive really the These early studies that show how structure affects electrical function Have grown and actually I didn't have time to To change this this slide here But Natalia Trianova just has just published a paper in nature communication that I think is also a milestone And it's worth to look at it was just published a few weeks ago the study shows how Models anatomical models obtained from MRI Can be used to predict sudden cardiac death in a cohort of patients So it's for the first time an image-based study of cardiac electrophysiology shows clinical potential This is a early studies that were published in the same Area so this one for example was using pig hearts with to show ventricular tachycardia Circles of ventricular tachycardia and then has been another paper by King's published recently on this Natalia's the recent one on nature communication is the first one showing the Potential of these MRI space models of cardiac electrophysiology to predict a clinical outcome So I think it's quite important and it opens quite a lot of doors for us in in I think just to To finish as part of my talk. I was told that we cannot have a break, but so we're just gonna have to do this this is this is quite quite Exciting for us and is studies where we're trying to understand phenotypes in hypertrophic cardiomyopathy Which is a disease that is inherited. So it has a genetic component And it's characterized by the thickening of the left ventricles so in some patients the ventricles are thicker and In addition to these thickening there is electrophysiological remodeling. So the properties of the ion channels are abnormal so the important feature of this of this Disease hypertrophic cardiomyopathy is that the phenotype is very heterogeneous So people having this disease can can show very different phenotypes and we wanted to investigate We have a database that includes electrocardiograms and MRI and we wanted to investigate whether we could discriminate Different phenotypes using just the QRS in the electrocardiogram So using just the first bit of the electrocardiogram that is determined by electrical activation The main underlying hypothesis here is that abnormalities in the structure of the ventricles would affect activation sequence More and that would be reflected in the QRS. So The different phenotypes could be determined by this QRS morphology This is a database the numbers have grown a bit now, but it's a rare disease. So it's not easy to get Patients we are getting more patients, but not with halters So the halters are 24 hour recording so they allow also Rate dynamics to be investigated and the study was done blindly on the ECG only So we didn't have all the clinical data on MRI and genotype data when we did the ECG analysis We investigated different features on the QRS so We did some QRS signal processing analysis to obtain Standard biomarkers like the QRS with is the most standard one and most people think that QRS would be longer in hypertrophic Adirmaopathy patients because the hearts are bigger and therefore it would take longer for the for the wave to propagate But in fact that doesn't seem to be the case and what then we did some because we geeks We did some mathematical modeling of the QRS and we approximated the QRS waveform through a Set of basics hermit functions and the coefficients would tell us how similar the QRS is to each of these bases so what we did then is one one is a Binary classification so based on these QRS features can we identify control from hypertrophic Adirmaopathy this was interesting, but actually not for the clinician So for us it was interesting can we do this can we differentiate healthy with respect to hypertrophic Adirmaopathy but the the clinician said we can do that already so that's not very interesting The most interesting bit is whether you can identify clusters of patients So can you discriminate? Patients that have a different phenotype Can you tell us whether there are patients that have a normal QRS an abnormal T wave that have a Very abnormal QRS and whether those clusters are related to clinical phenotypes at all So we did this The first thing is that the standard biomarkers didn't show any differences between control and HCM So this belief that the QRS with is longer in hypertrophic Adirmaopathy wasn't shown in the data There were people with hypertrophic Adirmaopathy with a normal QRS with and none of the other biomarkers Led us to any good classification between the two because there is overlap between healthy and HCM So we did this clustering analysis and we did see three clusters in the data and In this cluster here the cluster one that you see here the patients had a very normal QRS complex So they didn't seem to have any abnormalities in the activation sequence What whereas the cluster three were clearly people with very very abnormal abnormal QRS complex These ones were intermediate and What is more interesting is when we looked at the T wave the patients that had a very normal QRS Had a very abnormal repolarization sequence So not only the syndrome is the disease is quite heterogeneous. It's hard to understand the findings So we we looked at the differences in the clinical data between the three clusters Genotype didn't tell us anything. So we couldn't see any difference in genotype between the three clusters But the numbers are not enough So we want to increase the numbers of patients to actually understand a bit more what's going on the the cluster three Who are patients with a very abnormal phenotype had a very abnormal wall thickness and a higher ejection fraction So that was significant. But what was interesting is that the cluster one which and these are the patients who had a very normal QRS We're at higher risk of syncope so No more QRS high risk of syncope abnormal QRS Abnormal wolf wall thickness and ejection fraction. So we are trying to understand this a little bit more We did some Analysis of the ECG tried to understand these things So this is an example of a QRS of a patient in cluster three, which is very We called it wavy and the clinicians had fractionated. So it was an interesting exercise of of language so We we try to understand why why these things happen also in cluster three the the The abnormalities in the QRS are related to areas Indeed in the heart that that are areas that were had a normal Structure, so they didn't correspond to abnormal thickness. They corresponded to late activation Sequences, so there is a lot we don't understand on this the QRS abnormalities Cannot be explained by the location of the maximum LV hypertrophy in cardiac MRI They might be explained by other things and I have a hypothesis about it But we did we couldn't correlate with MRI wall thickness So what we are doing now and this is also a collaboration with Barcelona is to build patient specific Models of the anatomy of these patients based on the MRI with information on the ionic currents to surface body potentials in order to understand these QRS morphologies and Whether the modeling and simulation can help us explain these different phenotypes in the patients Because we cannot do it with the data themselves We need to build more complicated models that allow us to do this now the problem itself is Highly complicated the models are huge. They need high performance computing is just recently that we were able to Simulate the activation sequence in these patients I'm not sure. This is so this is this is way that Anna Mitchell and Ernesto Thakur did with a lot of people and It's based on an anatomy of one of the healthy volunteers And we were able to simulate the ECG for this patient for this healthy volunteer So I wouldn't say this is the ECG of this person What I would say is this is the ECG that comes out when we incorporate the anatomy of this patient, which is different So this is going to be a tool which we haven't done yet This is going to be a tool that is going to allow us to assess how the information that we obtained from the Patients in terms of the structure what's the anatomy of the ventricles with the fiber orientation? What are the abnormalities in fiber orientation? How much that can explain the electrical phenotype and the abnormalities in the QRS It's very exciting and it's certainly an area where you cannot do it without modeling and simulation And I think that's what makes it more interesting for me the idea of all of this and the vision behind it is also to use this type of Very complicated models for in silica trials for drugs and these devices So if we construct an atlas of models or a population of virtual hearts We can use them you know in order to test the efficacy or the safety of drugs And this is something we're doing at the single cell level already with a bunch of companies and with the FDA as well We are trying to evaluate how the computer models allow us to predict the safety or the efficacy of the treatment But we can take it to a bigger level where we incorporate not only single cell dynamics, but also Differences in anatomy in body mass index in structure in the models and we can go towards Multiscale models of the heart that allow us to test the efficacy and the safety of drugs and electrical therapy In silica before going in into clinical trials Okay, so this would be all preclinical testing but very useful so and we have now some doors open towards doing that this is an example of a simulation of the atria The propagation in the human atria as well and at the moment we have only one one anatomy But this doesn't mean we cannot create other anatomies and in fact given the the the big progress in in MRI space modeling, I think we can do it The electrophysiology is harder. I think So just to finish I think I think that the going from images and Data into constructing models from electrical decoding is something that is really really challenging and we kind of know how to do at the moment and I think it's a Really a promising area to be looking at the effect of mutations on the Different phenotypes we can look at the effect of structural defects on on on the function of the heart Many things that we have been doing very well. I think for many years now I think the main challenge for us and for biomedicine is to look at population heterogeneity Due to environmental factors and for that I mean anything that is external to the thing So if it's a cell it's what is regulating the cell that is not internal to it like a genetic defect when it's a body what is Regulating our behavior like what we eat circadian rhythms, whether it's day or night so the more there are more More and more papers for example that are showing that the maximum ionic Conductions that we all have been considering as parameters in our model are not permitted They are variables and they are variables that are regulated by what what we eat hormones So there is a paper showing how testosterone can increase the potassium channel by 30% and that actually protects men from arrhythmia So We in the models we have always been assuming that certain things are parameters and we believe it I mean we've been saying it to us ourselves for so many years that we believe that the number of ion channels in it In our cells is constant and it's not It's actually a way of cardiac cells of adapting to environmental factors if we eat sugar if our hormones change if It's day or night if temperature temperature changes or with if we take drugs The number of ion channels in our cells is going to change and it's going to help us adapt to those environmental factors that actually has a very strong influence on our phenotype and We don't know how to measure it and we are not going to be able to measure it and We are not going and we don't know how to deal with that huge uncertainty I mean some people call it uncertainty and they try to capture mathematically using Bayesian approaches and stuff for me It's just an unknown our our hearts our organs are open systems that are affected by environmental factors And we can measure things in the lab or in the clinic But this will change that the the next moment where the patient goes away from theater I went to the car to the cat lab the other day to look at an arrhythmia patient who was having an arrhythmia And it was very consistent and they wanted to ablate the patient went into the cat lab and the arrhythmia disappeared that happens all the time and that's just the proof of how much of an open system our organs are so as mathematically as mathematically minded people we want to capture things in in Equations and using parameters, but actually we need to understand that reality is very different and biological systems are Open and they are affected by these environmental factors. How to treat them. I don't know so the the way we deal with that with this in this study was brute force really so this is a these are recordings that are taking human hearts using a sock of electrodes so literally the surgeon and Clinicians Pierre Lambias and the target put a sock of electrodes around the heart of Patients that are going to go for surgery and they obtain electrocardic readings We used the Ohio Rudy model that I mentioned before in this study and what we did is To construct populations of models rather than a single model We use thousands of models with different parameter values to capture the data This was based on methodology that we published a few years ago So rather than using one model We used a thousands of models that had the same equations, but the parameter values were different And we only kept the models in red which were the models that were in range with the electrical recordings So we call this a population of models calibrated with the experimental data in this case In vivo clinical recordings the you can call it uncertainty in the parameters in vivo is huge So what we assume is that it can vary quite a lot and what we wanted to investigate is two phenotypes that were Observing vivo. So these are These are different cycle lengths or rates and what we observe is some patients had alternates which is bit-to-bit variations in APD in actual potential duration that were closing at high Pacing rates and other patients looked at how showed Falk type alternates that were consistently observed at high pacing rates So the first thing is that the models were able to replicate these two phenotypes and Then what we did is we investigated the ionic basis of these alternates and we looked at again What is the added value of these models? So we were able to replicate these two phenotypes and the added value of the modeling and simulation that cannot be done Clinically is to show what what was different between these two phenotypes and it was all very complicated Very complicated systems biology that only a person like Shinzo can do which was digging digging digging digging into the actual mechanisms of these calcium Dynamics and we show that these differences in alternates were actually Due to a difference in the L type calcium current What Shin did is to actually look for an anti-arithmic strategy in these patients for these patients that display alternates and Looked at how blocking the sodium calcium exchanger could lead to the disappearance of alternates now the the whole the whole Framework for this investigation was based on the fact that we don't know what the ionic currents are in those patients And we can never know So let's go for a modeling approach that explores a variety of mechanisms there That is absolutely wrong in claiming that these models represent those patients But that allows us to explore different possibilities and what if we block the sodium calcium exchanger what would happen here? so To me the the importance of experimental environmental factors and population heterogeneity and the many factors that affect our heart Calls from a different approach to modeling and simulation that is far away from Let's define what is the conductance of this particular cell at this particular time and explore things from a much freeway Using creative approaches there. There are a number of papers Calling for this type of approach and this one is quite interesting the universally sloppy parameter sensitivities in the insistence biology models And it's worth reading and then coming up with our own solutions What I think is also important is be tolerant because we are in a world of modeling and simulation Where we are exploring the use of modeling and simulation for different scenarios and there isn't one single solution We need to explore the use of modeling simulation in many different ways so one Considering models of us representations another one is considering models as tool and nobody can claim that we have discovered the best way of doing this It's a bit of a mess really so just just to finish and For me computer models are tools with her representation and instrument So it's ideal when you have the best possible tool for the for the thing you want to investigate but often actually choosing a Different one is going to do the job and you're going to get your publication faster and sometimes Timing is really really important. I don't mean not going for the ideal I think this is the absolute best but sometimes going with a tool that does the job can can be really fruitful and If somebody uses the wrong tool to wash a Glass and it breaks. It's not the tool that was wrong. It was the use of it. So for me using models Needs to be evaluated in the context of the simulation and experimental system And and we always need to think about models as tools with her representational instrument So in conclusion so computer models from from for me are tools for exploration and discovery and it is important to assess the success or failure of the model with respect to the discovery cardiac research is full of discoveries through iterations between computer model simulations and experiments in basic science and increasingly in translational and clinical settings and and I'm very happy to see studies that are reaching the clinical Journals the biggest challenge ahead is viability due due to environmental challenges And this is not just for modeling and simulation. It's for biomedicine in general and precision Precision medicine and personalized medicine. It's a hugely important problem And and we need to come up with creative approaches that help us explaining models as tools to explore the effect of this viability so there is increased complexity in what we do in modeling and simulation and we often collaborate in teams and and We are specializing in one area, but We work in interdisciplinary collaborations, which I think is very important in a Accelerating impact in biomedicine and it goes for I think from tolerance really so so from trying to understand the perspective of other people who come Who come from a different perspective and the typical one is clinicians and and engineers But there are many others and I was given a very useful tool for that Which is the principle of charity and it's when somebody tells you something always assume that they are very smart And they are coming from a very valid point and that actually makes the research much stronger in our setting and I think that's all from me and thanks for The very long talk. Thank you. Thank you very much for this splendid overview Let's maybe start a little bit provocative I think from some of the things that you said I seem to be able to conclude that Personalized models like a lot of people are trying to make like patient personalized models. You say I'm never going to work Well, I mean that was really not what I meant, but Because we're doing that as well. What I mean is it depends on what we claim So I think it depends on what we are trying to claim with them. So for example The study was referring to that Natalia Trianova has published now. They are working. So in fact, I Discussed this with her. So she's taking her study if you haven't seen it. So her study takes MRIs from patients that have myocardial infarction and she constructs an anatomical model of that heart with a scarf Personalized for each of the patients and then she runs simulation studies of arrhythmias and then she uses that to predict the risk of these patients to have Arrhythmias in reality so that she has a database for that So the models are personalized only in terms of the anatomy and a lot of people thought she would fail Because there is no way that if you don't personalize for the function for the electric for the electrophysiology This would work, but they are working So my question to her was all has always been so what do we learn from this? So if you personalize only for the scar and the anatomy and you are able to predict risk in that cohort of patients What do we learn? One is that the electrophysiology in those patients is not critical for the evidence. So For me, it's great that it works. It won't work for all the patients and that's where I come with Tolerance, I'm not going to come here and say they're never going to work because I think they are and they are working Even we we didn't anticipate that it didn't so I think for me It's great that she's showing that it can work in that setting but I also think there are other ways of doing things that are equally valid and What I don't want is people to think that patient specific modeling is the only way to go Because it's not because for environmental factors, for example, it would be quite challenging. So for me, it's Let's explore what modeling and simulation can do so patient specific modeling Yeah, go for it show that it's going to work But there are other approaches that don't take that turn that can be equally valid for phenotyping Yeah No, I totally agree with you The only thing is that sometimes reviewers don't agree that if you come with a simpler model that answer the questions They say like why didn't you do a 3d personalized model in order to do it? So still there's work to go there. Yeah, but also I think that's also educating reviewers not to be a pain in the arse I think that's impossible unfortunately. Well, I mean, let's be critically constructive right especially in this so I sit in a panel for funding and in the welcome chest and The people who really know how to do this constructive criticism is neuroscientist Neuroscience is a community of people who criticize very constructively They are very friendly and they always get the funding and other communities are Perhaps less mature in terms of interdisciplinarity or in terms of how we review papers and applications And you can clearly see that different communities have different takes of the on thing I think neuroscientists are doing great So we can learn from them. Yeah now another question that you partially Touched also and that's a thing that we for example are struggling with also and I know several people are struggling with It's what you say is indeed is variability The what do you call it uncertainty or variability or changes with environmental factors and things like that But the question is how do we deal with it? You showed like one is brute force The other thing is some people are claiming like maybe in our mathematics we have to try to integrate it from the beginning or Do maybe some kind of machine learning or there is like what's your view on all these different kind of approaches? I mean what what I think is that all all the methods are need to be tested. So so this is a huge problem and So for me uncertainty quantification and viability are different things. So one in terms for example on how we do uncertainty quantification for Determining the conductivities in the tissue so Conductivities in the tissue come from structure that is going to change very slowly probably So we can consider that as a as a parameter because it the changes are going to happen with years right age and So I think using Bayesian approaches to estimate uncertainties coming from the data and Identify uncertainty and quantify uncertainty in that setting makes sense. I Cannot anticipate whether somebody is going to have a very very clever idea to do things in a different way in the future And they will come up with a better solution and I'm looking forward to it Conductancies of iron channels are a different problem Because they cannot be measured and they change really rapidly and In a matter of a day they can change by 30 percent So and that's what we know of what happens in a dish they cannot be measured because The the the electrical electrical recordings are done in isolated cells that are damaged So we don't know what they are in tissue and when we take them from the heart So to me I only currents are about viability not uncertainty. I mean there is uncertainty But the huge problem is not uncertainty Viability is they they are varying now as we speak, you know your channels are changing And they are being affected by day at night temperature what we eat so For me that is a different problem and and it calls from a different approach and we because we are dealing with With variables that are not parameters. So you can try to to use Bayesian approaches for that problem but I think The assumptions you're making in this in the use of Bayesian approaches to investigate uncertainty are so huge and so not supported by biology or experiments that perhaps Other approaches are going to be more useful, but I think it's worth exploring all of them. I mean there is no No reason for not going for all of them and trying to see What is more useful in what setting the the intolerant view that it's needs to be one way I think damage is science and What about measurement inaccuracy because that's another one that you have would you treat it differently from variability or Is it better to include it together? We include it together because we have no means of differentiating so We just yeah, I mean it's huge problems, but The measurements are the ones we have that there are also different ways of of dealing with viability. So the So people in the US Eric Sabi and David Cristini rather than using the population of models that we are doing are using a different approach Which is coming up with? algorithm so coming coming up with Stimulation protocols in cells that would allow you to estimate the parameters more Robustly so rather than that doing traditional electrophysiology in the lab doing much more sophisticated electrical measurements in cells In order to obtain Models for specific cells and I think they want to control the experimental error as well So what they do is they they do very extensive Stimulation protocols in a limited number of cells and they come up with models that are specific for those stem cells Now, I think that's really important where They can only do that in animal cells because the human tissue is Not going to be Given for that. So if people have human tissue for their own experiments, it's very rare because it's so Not available. People won't give it to somebody else to do their experiments to construct a model specific model So they they use guinea pigs or use that they use and They can control experimental over there. I can control things in a much better way But when when you talk about human experiments You you are given what you're given from your collaborators You're very restricted. So you need to deal with the especially if it's in vivo How many variables are affecting the human heart the error? How do you estimate that? It's it's not under your control. Even if you have a very good relationship with that They're partially related to that It's like you have a lot of experience with cellular work and things like that Which is mainly done in a really controlled way But we also have been working with clinical kind of things the data that you get from clinicians is totally different Both in quantity and quality than what you get from biologists or cellular electrophysiologist Do you change your approaches based on that or do you say like okay? This is what we have and we try the same approach So we change the approach here and we try to do our best. So Yeah, of course, we change our approach and and we but but it So reality is reality. Okay, so reality comes from you you are given what you are given Clinically you cannot do more So if Pierre and Peter are putting a sock around the heart of certain patients and we get those recordings from Ikele Oriini, actually So we do what we so we are grateful to the patient that has allowed us to do that We don't go and ask. Oh, actually, can you put this drug on the patient? So you can you are working in a setting that is not mathematics. It's reality so it's reality of people who undergo a surgery its reality of the Recordings that are done in an invasive way. Those hearts are not in the body. They are It's open chest. So I think in doing that in design in the modeling and simulation study It's very important to take into account what reality is So when a pure mathematician comes and says, oh, I would use Bayesian approaches because that's the way I say, okay Let me let me tell you my story So it's we are hybrids. We're not any more pure engineers or mathematicians any other questions coffee time Good. Thank you again. Thank you very much