 So, we are talking about the Virtual Brain Assimilation Platform and I think you know why you are here, you know the platform, you know that it is open source. It started publicly in 2012 and we counted our downloads and it's now over 21,000 which is pretty nice. So, the Virtual Brain is also part of the Human Brain Project and of several other larger consortia. What does the platform offer? In one sentence it integrates or links neuroimaging and computational modeling. So, what it does is it provides you a tool to construct personalized or individual brain network models and to simulate individual brains on the computer. And this is actually Jessica, our colleague who is not here today but with her brain. I show you the pipeline that we have developed that basically takes into account all the different imaging modalities. We reconstruct the cortical surface, the fibers in the brain from diffusion tensor imaging. Then we passulate the brain in different regions. There are different passillation schemes available and each region then is represented by a population model and they are interacting through the anatomical skeleton. So, this brain model then simulates real activity. I have to apologize, I have also bad cough. So, you simulate activity like local field potentials, population spikes, you can also simulate EEG, MEG or FMRI signals. So, the imaging pipeline that I just have shown you is available publicly. Some of you I think, at least one of you has used it already, Jan. So, you can either go to the GitHub or you can also use pre-installed versions on different supercomputers. One is running on the neuroscience gateway cluster and another version recently with some still limited functionality. I think it's only the structural connector that you can extract presently runs on the collaboratory of the human brain project. Who of you has access to the collaboratory? Raise your hand, please. Okay, quite many. So, many of the notebooks and the tools that you see today, you will find there and can use there directly. So, you might know already that we are not simulating individual neurons, we are simulating populations of neurons and why can we do this? Why is it not necessary to simulate each individual neuron? Because in nature and also in the brain we have the phenomenon of synchronization. And this is one nice example, these are memorations of starlings and you see how they are synchronizing to their neighbors. This is not as nice, but this is in Berlin, it's a TV tower and these are crows and no starlings and here these are no birds, these are neurons in the state space. This is the membrane potential, this is the opening probability of the ion channels and you see how the neurons are synchronizing and what you not see is the red cross in the middle, this is a mean field approximation of this population of neurons and mean field approximations are common methods and statistical physics to describe and reduce manner complex systems and this is what we are doing here. So again, the standard version that you can download presently of TVB does not simulate individual neurons, it simulates population models that describe the activity of groups, of large groups of neurons and as you can see here this is one example of such a population model and TVB we have about 15 different types of population models that you can choose. So here in this example we have a mean field approximation of the excitatory population of an inhibitory population and then the properties of some parameters at the cellular level average across the population and in this case we have the coupling strength, the excitatory coupling strength of NMDA receptors and the inhibitory coupling strength of GABA-ERJEC neural transmission. So now each of the nodes is represented by such a population model and they are interacting through the individual anatomical skeleton. Here you have the evolution equation, sorry, that computes now the trajectory of a region through the state space. So what you see here, hopefully you can recognize it, it's a little bit light, the differential operator on the left-hand side, on the right-hand side you'll see the neural mass model N and the vector of state variables of the specific neural mass model which can be local field potentials or the mean field synaptic activities or population rates. So now here you see the first sum operator, this sums the activity of the connected nodes J and in order to weight the sum of individual activities of these nodes we have the structural connectivity matrix that contains typically inferred from diffusion tensor imaging tractography data the connection weights between the individual region pairs. G is a scaling factor because the connection strength that we obtain from tractography are relative values and they need to get rescaled in order to use them in a meaningful manner. The time courses or state variables are summed up with time delays that are given by D and depend on the connection length or distance between individual region pairs. So here this is some optional part where you can also add local connectivity, I don't know whether we have a use case of surface simulations today. No, I think since you never did it before we just leave that out. It's also possible to simulate the activity at far more detail taking into account the geometry of the individual cortical surface and for this we have a local connectivity kernel, we have also a scaling factor and we have a theory time delays that I think typically are not used in the local connectivity. And here this is interesting again for us. This is another input current that can mimic stimulation if you want to add stimuli to brain regions and see what the evolving dynamics would be and this is a noise term that you can add. So now where are the personalizable parts of our model? I just highlight them here. So we have the usually individually fitted global coupling scaling factor G that rescales the structural connectome. Of course the structural connectome can be also individual and the time delays are inferred from the distances that you get from the tractography between the different brain regions and if you want to add the surface geometry, this would be also another individual aspect of the brain. Now I would today also want to always give you a hint what the latest developments are and the kind of cutting edge research ongoing with TVB. So therefore I would like to mention that we are presenting also at the INCF booth here at the INCF conference a development where we do core simulation, multi-scale simulations. This means that most of the regions and the brain are represented by population models as I just described. But some of the regions are represented by more details. For example microscopic cellular network models. And this is one example we are working on right now where we have this microscopic detail in the hippocampus and the entorhinal cortex and in the dopaminergic midbrain. And this detailed connectivity comes from the literature. So this is nothing that you can infer from diffusion tensor imaging data. This is really manually selected from the literature. Ideally and this is also parallel development in the human brain project. We can in the future get this information from a platform, from a library or from atlases that are readily available. Presently it is quite painful to get this information. So now you can select some of the regions at more detail and the question is with which tool do you want to simulate the microscopic networks. We in our first proof of concept choose Nest as core simulator. Nest is a very far developed simulation platform that has many nice features that in part TVB has not. For example, multi-threading. So you can run your simulations on different processes. Also you can run the simulations on supercomputers. It has many different types of micro-computers. You can run microscopic models, synaptic models, circuit models and neural network models. So this is another notebook that you can find on the HPP Collaboratory if you are already locked in. If you go on the neural informatics platform, you can go in the menu item tools. And there I think at the top, because we just edited it, you will find the core simulation Jupyter notebook for parallel simulation with TVB and Nest. So this is now a traditional simulation where we only simulate at one scale, the population scale. And here you see how we simulate 20 minutes of resting state FMRI activity. And at the same time, while we are producing this FMRI activity, at each of the individual regions we can infer the underlying neuronal interactions. What we are plotting here are the firing rates of the excitatory population and of the inhibitory population. And we now see something that we cannot measure typically. So we really now can infer the balance of excitation and inhibition on an ongoing level. And of course, if you wanted to measure this in an individual, you would have either to stick electrodes on the brain, but also quite many. So in the normal case, that would be not possible. Here we in the simulation use a reduced one-run model, another example of a population model. You see the input currents that drive the excitatory and inhibitory population. They translate into firing rates for the excitatory and inhibitory populations. This is what we are plotting here. And here is a synaptic activity for both populations. So if you are interested in the studies, this is published in E-Life. This is a very full figure. And I think you will not recognize everything, but what it shall tell you is that with a single model, with a single parameter set, it is possible to capture many different temporal and also spatial scales. So what we see here are the features of our simulated data of a single model with a single parameter set. And here are corresponding features of empirical studies. And you might see that the features of the empirical studies come from monkeys, from rats, from monkeys again and humans. So they have totally different sources. Some are recorded, some of the underlying data are recorded invasively with electrodes in the monkeys and in the rats, others with non-invasive imaging tools. So the nice thing about the virtual brain, and this is also at the heart, and this is the purpose of the virtual brain is to integrate these different snippets of observations in a self-consistent model. So for example, here we see the relation of firing rates. So these are firing rates and the phase of the membrane potential. And there's a certain relation that has also been shown in empirical data. Then here you see the inhibitory and the excitatory input currents that drive neuronal populations and they have a certain amplitude ratio. And this ratio has been also published several times in empirical data. Then here you see the relation between firing rates again and the power of the field potential. And again, this is something that has been observed and published in empirical data. Here this is a relation between the EEG, Grosem power, the alpha power and the bold signal, which has an inverse relation many times published in the literature. We get also some features of complex systems like scale freeness has been published for resting state of my data, also for EEG data and the switching of the functional connectivity, the functional connectivity dynamics. So all these different features are integrated in a single model. And we now can manipulate the model and see if we change a certain feature in one scale. How would that impact the behaviors at the other scale? So this is how we really mechanistically understand how these observations are linked together. So this is again the simulation. You might have been surprised that we simulate more than 20 minutes of resting state FMRI. And if you know complex systems, if you have no further information, it's almost impossible to predict 20 minutes of a complex system in advance. So we had a certain trick here. We call it hybrid model. We injected EEG source activity. So at every time point, we provided a little bit more additional information about the state of the system or one component of the system. In this case, we use the EEG source activity to approximate the local excitatory input currents. And we replace this term in our model by the EEG source activity. So with this, the model is not really autonomous anymore, but it has the advantage that it can be very facefully, it can very facefully simulate the resting state bold dynamics. And as I mentioned already, we see the underlying local and global interactions that give rise to those observations. So we have now a hypothesis how our observations are computed by the brain. We see here the FMRI signal, the bold resting state networks, and here the relation of the simulated bold signal to the EEG alpha rhythm power. So now I show you one very new development and I show it today because I have been asked in every node workshop about it and we are working on it and I wanted to share it with you. So what is always of interest when one hears about the virtual brain is does it really simulate function? And then usually we have to say no, it more or less simulates the dynamics that we can observe, but they're not really function. So what we know is that functional connectivity is related to intelligence. This has been published a couple of times. One publication is based on the human connectome project data that we reanalyzed here and we reestablished in this analysis this already published relationship. The functional connectivity strength relates to the score of the participating individuals intelligence score. Interestingly, the response time also correlates with the intelligence score of the individuals and if you group now the participants according to their response time and look at the relation between the response time and the function connectivity then also there is a very clear and significant relation. So those who are more intelligent according to the score take more time, longer time to provide responses in the task that was used to compute the intelligence score and they have higher functional connectivity in their bold data. That's interesting and the virtual brain actually is exactly the tool that could provide us now the means to explore what could be the mechanism behind this observation. So first of all we have to make sure that we really facefully capture the individual differences of functional connectivity that we find for the different intelligence groups by our simulations. And for this we advanced the standard model so typically for the long range connectivity we only have input of the excitatory population to the distant excitatory population. What we added now is an input of the excitatory population to distant inhibitory populations. So this is what you see here. These are the excitatory populations. This is a traditional connectivity scheme. They connect to each other, they excite each other, but they only excite their local inhibitory counterparts but have no connection to distant inhibitory populations. So these feed-forward inhibitions which exist in the brain, everybody knows that, are now added to the computer model. And the ratio between long range excitation and feed-forward inhibition is now optimized so that the fit between the functional, the empirical functional connectivity and the simulated functional connectivity is maximized. And by this we get very good predictions of functional connectivity and we can capture the differences of functional connectivity that we observe for different intelligence scores. So now we have the models and we can relate the state variables of the model to functional connectivity, for example. So this is what we have done here. We have looked how the input current in our model relates to the functional connectivity. And what we found is, again, functional connectivity is higher with higher intelligence and with higher functional connectivity and higher intelligence, the input currents in the populations decrease. So the input currents, again, this is the reduced Wang-Wang model, this is the equation, this is a drive that goes in a single population. And this, on average, decreases in more intelligent individuals with higher functional connectivity. What we also found was that the correlations of these input currents are higher. So this higher functional connectivity, this higher functional connectivity in two or in one region pair in the bold data, then also the correlation of the input currents would be higher for these regions. So now we have our model and we want to see how does the functional connectivity, the large-scale network dynamics interact with cognitive function. And in order to do this, we now add some functioning functional circuits in our model. Here you see an example. This is capable to excited-tale population, mutually inhibiting well, inhibitor population to simulate working memory or to mimic working memory and decision-making. And I will explain to you in a moment how this goes. But basically, for working memory, you simulate attractor states. This is what we present, the maintenance of the memory or for decision-making, you have a winner-ticket or competition between these two excited-tale population that are mutually inhibiting. And we place them in the prefrontal cortex and the lateral-entraparietal area, the areas that are known to be related to working memory and decision-making. So now for working memory, what we do is we inject a stimulus at a certain time point and one of the two excited-tale populations, then if the stimulus is strong enough, an attractor state is achieved and then after, in this case, two seconds, a second stimulus is applied. And this can now either be too weak and the attractor state of the first population is maintained, then the memory would be robust. Or the second, distractor stimulus would disrupt the activity, the attractor state of the first population. Then the memory state would be disrupted. Or we have also the scenario that no memory trace has been formed from the beginning of the stimulus was too weak and no memory was elicited. So now we can look at again the two conditions, low intelligence group, high intelligence group. We noted already in the low intelligence group, the input currents that go on the populations are high. In the high intelligence group, the input currents that go on the populations are low. And these graphs now show you the red part. As you can see here is the robust part where the attractor state is maintained despite the distracting stimulus. The blue part is the one where the memory is getting disrupted with a distracting stimulus. And the white part indicates that no memory trace has been shaped or formed. So what we see now is that for the low intelligence group, only a small distractor stimulus already disrupts the memory trace. While in the high intelligence group, you need a stronger stimulus to disrupt the memory trace. So the memory and the high intelligence group is more robust. On the other hand, you also see that for the high intelligence group, you need from the beginning a stronger stimulus to induce an attractor state, a memory state. So this could be beneficial as well. Because this means that only small stimuli would not elicit a memory and the resources would stay free for more salient or prominent stimuli. So now I come to the decision making simulation. Again, we have the two excitatory populations that are cross inhibiting. And we now are inspired by a visual stimulation paradigm with random dot motion detection. So this random dot motion has to get interpreted. And basically both populations are stimulated. One population gets a little bit more stimuli or evidence than the other. And this difference between the input and the two populations is called contrast. So now we see what is happening when we run a simulation. Again, for the low intelligence group and the high intelligence group, what we see here is a firing rate. We see here the stimulation period in blue. And we see already very quickly that for the low intelligence group, one population is ramping up its activity very soon after the stimulus started. So when a certain threshold then is reached, then a decision is taken. So basically, the population would be caught in a basin of attraction. You see here for the high intelligence group that the time of integration is much longer. So there's a less deep increase of activity in the blue population. There's more time for accumulation of this ambiguous evidence. And this means there's more time to average out all the noise and averaging also the evidence that then leads to the right decision. And this is what what you see here at the bottom panels. Maybe first I focus on this one, you see the dependency of the integration times of the blue line from the input current. So again, high input current, low intelligence, low input current, high intelligence, the higher the intelligence, the longer is the integration time. And this is what we also found in the empirical data. So if we now look not at the input current, but the input correlation, the input correlation is related tightly to the functional connectivity and the higher the input correlation, the highest intelligence. And what we see here, here on this axis, you see the contrast of the stimulus. If we pick an intermediate contrast, then we see that this higher input correlation, we go in a more wide area and more wide means that the decision performance is better. So we see that the network dynamics really influence decision making and also working memory performance. So and we reproduce what we have seen in empirical data, the relationship between the mean reaction time and the functional connectivity. But we want to understand the laws, the rules behind those on the real mechanisms, and not only see correlations. So for this, we look now at the state space. So we have the face portrays here for contrast 0.2%. This means that the population A, and this is a state variable of population A, this is a state variable of population B, receives slightly more evidence and input. So the correct decision would be for population A, or direction A. And so if you do not understand the diagrams, you will hear a lot more about this in the next session, for example. So this is just the first circle of the learning spiral. But what you see here are the flow fields and should be purple, looks more gray here. You see 1000 simulations. These are all the gray lines in each of the panels. And the orange are the average of the 1000 simulations. And we did that for different background input correlations. So here, I just call it functional connectivity, but it's the input currents, it would be no correlation, 0.5, 0.9, and 1. And we now can look how this influences the behavior of the system. And what we see is that, so these are two basins of attraction, the full circles, that for higher background correlation, the state variables stay closer to the so-called separatrix, the red line that separates the two basins of attraction. So we see here both deviate very soon with low correlation and are attracted to one of the two basins of attraction. While it is high correlation, you see that the state variables stay much closer to the separatrix. And this is good because if they move away soon from the separatrix, then as you can see by the flow fields, the velocity of the flow is getting stronger and it's getting more and more difficult to move the state back to the other side. So if they arrive at a certain point, then even the following subsequent stimuli would not be able to pull it back and it could come to a wrong decision. If there's more time for evidence accumulation, if the system stays more balanced for longer time between the two states, then corrections are still possible. And as I mentioned already, you have evidence integration, you average out the noise. And as you can see here, you come up if you have a very high correlation with the right response. So this is just an example how we can really use the tools that are available for nonlinear dynamical systems in order to understand the mechanisms and even related to cognitive performance. Now I switch topic. So this was the healthy brain. Now I should give you some examples how we can use the virtual brain for clinical application. And this is an example that comes from our colleagues in California, Urbine, actually they moved now to Texas, Dallas. This is in the graphical user interface. So now you see here for the first time the graphical user interface of TVB, a stroke brain with a big lesion. And they had several of those. So each of those dots represents one stroke brain. So they virtualized them and they inferred a couple of model parameters for each individual by fitting the simulated functional connectivity to the empirical functional connectivity of these individuals. And then they looked whether the parameters that they fitted are predictive to the functional outcome. And this is a functional outcome. The motor assessment Fugl-Meier score, it's a clinical score that basically rates the degree of motor function of these individuals. I think the higher the better. And they found that several of their parameters including this, I just selected one here for illustration, the coupling between the local coupling between the excitatory and the individual populations, were predictive for the outcome of the rehabilitation six months later. And if you want to look it up in their publication one year later I think the prediction and accuracy was even higher. So now I give a short preview for the work in the field of epilepsy by our Marseille partners. You will learn more later by Julie Couture who actually joined our team in Berlin from Marseille. And here you see a virtualized patient brain with intracranial electrodes. So this is something that is done in these patients. We stick electrodes in the brain in order to find in patients with epilepsy the origin of the seizures and also the distribution pathways. This has to be done because on 30% of the cases the drug treatment does not help and only neurosurgery is an option to help those patients. The neurosurgeon needs to know where the activity, the pathological activity starts and how it propagates through the network. So here you see how in one hemisphere some pathological activity starts, it propagates through other regions and then terminates in the other temporal loop. So this is another example of a patient and here you see how the clinical team and the virtual brain bills determine the origin of the seizure of the epileptogenic zone. I think the red is the clinical team, the green is the computer prediction and the yellow is the overlap. Since one patient they looked at several patients and then after surgery has been done they looked at the outcome of the surgery and the surgery was done in accordance to the predictions of the human clinical expert team, not of the computer. So here we have again a clinical score, the ankle score, it's actually on this x-axis. Three means there was no improvement for these individuals, two means for those there was a slight decrease of seizure frequency and for those there were no seizures anymore. Here on the y-axis we see the divergence of the predictions between the computer model and the clinical expert team and the highest is divergence, the worst was the outcome and this was significant. So the computer model could predict the outcome for an individual after surgery and this was sufficient evidence that the computer predictions might be meaningful to start clinical trials that started this year with 400 individuals at several centers in France and it's ongoing for the next four years and there in half of the patients the predictions of the computer model will be also used in addition to the clinical assessment for the planning of the surgery and then afterwards we will see whether the group with the additional advice by the computer will be better after surgery. So this is another application from our colleagues in Gantt Daniel Marinazzo's group where they virtualized the brains of tumor patients, brain tumor patients and in their publication they found that these local parameters and in particular the feedback inhibition can differentiate well between tumor-free and non-tumor-free tissue which might be of interest for neurosurgeons as well and here this is another example of a clinical application where our colleagues in Toronto in collaboration with several other colleagues used the data from the Sydney memory and aging study 16 patients with Alzheimer's disease I think 32 or so with mild cognitive impairment and more than 70 healthy controls here each dot represents either healthy control, Alzheimer's or mild cognitive impairment and it has been shown that the biophysical parameters of the model can make pretty good predictions of the cognitive scores but even nicer one can then relate the change of cognition to the different changes of the individual biophysically interpretable parameters so there's a claim that the parameters in the model have some biological foundation grounding and that we can use this to identify possible mechanisms and target points for intervention. Of course these are all very small cohorts on sample sizes and for clinical application we need validation of all these findings with bigger and different data sets so what is also a need is many diseases have a molecular or genetic origin to make the bridge between the molecules and the genes and the large-scale networks and I'll give you an example how we do this this has been the example I give you has just been published it's also about dementia so in dementia patients one sees an aggregation of proteins on the brain one of these proteins is amyloid beta another one is tau protein and for amyloid beta it has been shown in cellular studies and animal studies that in the presence of amyloid beta the inhibitory interneuromans my my function so here you see that the inhibitory postsynaptic potentials are smaller and also the firing rates are smaller in the vicinity of better amyloid plugs so there is a hypothesis that the function of inhibitory interneuromans is disturbed in these individuals so we can now use another imaging modality namely PET in order to integrate this additional information about proteins in our brain models so in the Alzheimer disease neuroimaging study acne we have at least partially for subgroups of patients pet data positron emission topography data that measure the burden of better amyloid on tau protein this is one example of a distribution of better amyloid average across the different region of the postulation use in this case what we do now is we link the burden of the better amyloid to the population activity and for this we have to create a new transfer model that establishes this link there we take the information from the literature and just implemented in our model so we know that with the aggregation of better amyloid there is disturbed inhibition this causes disinhibition and higher excitation so what we need now is a population model where we can implement this impact on the inhibitory interneuromans and one population model that is pretty detailed as the answer model consists of an inhibitory interneuron so it's part the pyramidal neuron and an excitatory interneuron and because it has sufficient degree of detail we can now alter the some of the kinetics parameters of the inhibitory interneurons depending on the better amyloid burden that we get from our pet data and this is what we have done and when we do this for all the available individuals again Alzheimer's disease mild cognitive impairment and high C control then we see just by incorporating and otherwise identical models the better burden and modeling it via the effect or cause and effect relations that we have established we get a slowing in the simulated EEG frequencies so we have here the frequencies and we see that in the Alzheimer group we have a peak at very slow frequencies in the set average which is not typical for healthy individuals and this is what you see in patients so here you see an example how we can use bifurcation diagrams to assess the effect of changes of parameters in the model so here what has been done is depending on the better burden the parameter has been changed and the inhibitory interneuron kinetics tau the time constant and this leads to either an alpha cycle so a limit cycle with an alpha frequency or two limit cycles with two different frequencies alpha and ceta or if we simulate very high burden of better amulets through modulating of this tau parameter we get a single limit cycle with a ceta frequency and if we look at the time series would be different patterns this would be an alpha reason alpha ceta reason ceta reason so this just gives you an example how one can use such tools like bifurcation diagrams to see how if you change a certain parameter of the system that dynamics would also change and mimic potentially certain behaviors that we see in pathologies this is an example of a virtual drug therapy where we use nemantine which is an NMDA antagonist used in dementia and we simulated by blocking the coupling between the excitatory interneurons to the pyramidal cells here you see the effect these are the frequencies this is a scaling factor g and you see again the groups alzheimer's disease mild cognitive impairment healthy control and the alzheimer disease has very low frequencies as you can see here if we just decrease the excitation by the excitatory interneurons to the pyramidal cells by making the effect of nemantine this can be in theory normalized and this is an example of a receiver operating curve so why are we doing all this what is the purpose our ultimate goal is to identify features in our simulations that are closer to the real mechanisms of neuronal interactions and better predictive for certain disease trajectories and how would we assess this so biomarkers are assessed by certain statistical properties or characteristics two of them are sensitivity and specificity specificity and usually your compute such a receiver operating curve and you have to be in this range in order depending of course on the purpose but for many purposes to have a useful biomarker so you have to maximize the curve and in our initial simulations at least we could slightly increase the area under the curve here it is working on that actually and I think you are doing already much better so this is an old result so we have already better improvement so this is at the heart of a new consortium that has been started in December last year virtual brain cloud where we focus on this work on neurodegenerative disease with a virtual brain we have 17 partners in Europe so it's a really big consortium and there are the main focuses to identify such computational modeling and third predictive biomarker for disease trajectories okay now I finish and I just show you some fun developments that we are also doing we are using the virtual brain this is not animated for public outreach projects this is the latest one this is an atlas that is now also available online that is part of a traveling exhibition it started in Israel, Jerusalem recently and travel through Europe where actually more at a level of students or even children the brain is explained and they can be experienced so it's experienceable by a touchscreen and the museums are big touchscreens and people can rotate it and zoom in and select regions and also get information about certain functional networks and have this in several different languages for example also in Hebrew and Arabic so this is easily feasible such type of work given that we have the computer models available and it's a very so it elicits actually very high interest which is nice this is I think the last thing that I should use is our art science work where we use new feedback and coordination with our app that we are developing and then this big event up to 20 people can actually use these headsets new headsets and interact so with their connected brains can create certain dream scenarios together so this is one example of one show that happened once in Toronto all the visual scenes and also the sound which we do not hear is based on some libraries that are controls by the activities of the 20 brains that are participating there and that's I think it thank you very much