 So if we look at brain connectivity, what you see here is a DTI-based structural connectivity, we can first look at different features of the system. Many features were already mentioned by Martin in his talk, so we can identify which club networks, we can identify different modules in the system, so you have a higher connectivity within modules than between modules. Another feature which hasn't been mentioned yet is that there is asymmetric connectivity. So it's something we can't measure in diffusion-transfer imaging. For diffusion-transfer imaging, you see a connection between A and B, but you don't get information whether you have an imbalance, whether direction A to B is stronger than B to A. But if we look at track tracing data in the food fly in the mouse or in the macaque monkey, we can actually see asymmetric connectivity that one fiber tracked or in the food fly one bundle of axons is a lot stronger than in the opposite direction and in some cases it's even absent. So fiber tracks or axons that only occur in one way of the connectivity. We can use this information to inform how we diagnose and treat brain diseases. So the idea is to use brain connectivity information to diagnose what subtype of disease is present and then to decide how to treat the patient, where treatment could be deciding what drug to give to the patient or what part of the brain to change, what part of the network to change. Change could be surgical removal for epilepsy, for example, or it could be invasive or non-invasive treatment. Invasive treatments we're interested in as an implantable device using optogenetics, that's one project. Another treatment I'm very interested in in the future is using focused ultrasound to change connectivity in the system. Because one problem with using drugs is that they're targeting the whole brain. Once the drug is going through the blood brain barrier, it's affecting all of the brain, which is leading to large side effects, compared to other approach that are targeted to small parts of the brain. So what do we need in order to reach that goal? First of all, we need a diagnosis for individual patients that includes etiology, so developmental origin, and also disease subtype. So we need to understand what was going wrong in the past to lead to a certain state we observe at the present, because this gives us more information about the underlying cause of the disease and how to treat it. So it's not just looking at the present state, but also having a prediction what happened in the past. We also need to identify potential treatment targets, which parts of the brain should be treated, and we need models for effects and side effects of treatments. So starting with the diagnosis, we can start to look at different subtypes of a disease, and the example here from our lab is looking at subtypes of dementia. We can start with the group level, where we look at functional connectivity at this stage, and we see two distinct changes if we look at the group of Alzheimer's disease and Lewy body dementia. So Alzheimer's disease is more frequent type of dementia, Lewy body aware type of dementia, but at early stages symptoms are quite similar, so it's difficult to distinguish what type of dementia is happening. But if we look at functional connectivity, we see two differences at the group level. One is looking at short and middle range connections. So we just measure for functionally connected regions. Are they nearby in terms of Euclidean distance between them? Are they at a medium range? Are they far away between each other? And this ratio between short and middle is different between those two patient groups. We also see a change in local efficiency. So this tells you how well neighbors of a node are connected, how easily you can travel at this local level, and this local efficiency is different for some nodes of the network. So this is just a result at the group level, but what does it tell us about individual patients in the hospital? We're starting by simply plotting those two features in 2D space in the scatter plot here. So you have on one axis local efficiency, on the other one the ratio between short and middle range connections. And you can basically just draw a line, a blue line in this space here. On one side you have all Alzheimer's disease patients in red, and on the other side you have almost all Lewy body dementia patients. So it seems to be possible to distinguish both groups by looking at those functional connectivity features. So this is one example where just looking at brain connectivity already seems to help. Unfortunately this is not always the case. As we already heard in the previous talk by Peter, you might have very similar connections but very different dynamics. So because brains are non-linear systems, a very small change in the network might lead to a very large change in dynamics. So we might have those different structures of the connect home on the left. And in one case the patient shows seizure, so you would say this is epilepsy. In another case the patient shows hallucinations, so you might say it could be schizophrenia. But the point is that you might have on the top the blue network which has very different structure to the network at the bottom, but you have the same phenotype, you have the same kind of behavior of the network. On the other hand you could have networks that look relatively similar, so the two at the bottom in red, but they show very different dynamics, very different behavior. So if you just have machine learning based on connectivity you would say that the patients at the bottom are in the same cluster, they're in the same group because they have similar connectivity. You only really find out that there was a difference if you look at the dynamics of the system, if you simulate dynamics and activity in the network. So we're interested in simulating dynamics and also simulating the development of diseases. In terms of development the idea is to look at early stages of brain development and then to simulate in the computer what happens over time. So starting from before birth or early stages to change connectivity, to change cortical thickness and to see what kind of changes lead to healthy development at the top and what kind of changes lead to some kind of pathological development, epilepsy, schizophrenia, Tourette syndrome and other diseases. So what parameters do we have to change? Are there some genetic parameters we have to change? Is it something about the timing of different developmental events that we have to change? And finally to use this information to predict how to treat patients. Starting with healthy brains we looked at data from healthy subjects just observing which connections are changing over time. For that we looked at diffusion tensor imaging data, structural connectivity in 121 healthy subjects age four years to 40 years and those are just snapshots. So unfortunately we don't have longitudinal data over 30 years so that's not available but there are snapshots for different ages and we can then test which connections, which fiber tracks are changing over time in this time range. In dark gray or light gray you see fiber tracks that are not changing in that range so they stay stable over time and this is basically normalized by the total increase in brain volume and total number of fiber tracks. So of course there is a change over time but if you normalize that in relative terms the strength of connections is stable over time. But you see that some connections show changes in where you see a decrease in the number of streamlines relative decrease. In blue you see an increase and this doesn't necessarily mean that you suddenly have new axons forming. It could just be that myelination is increasing over time and therefore you can detect streamlines, you can detect fibers that you can't detect during earlier stages of development. In yellow you find some fiber tracks that are gender specific so they might decrease in females and increase in males or the other way around. If you look at this image you already see that the fiber tracks and red are not just randomly distributed in the system. They're more frequent within a hemisphere so within the left hemisphere and within the right hemisphere. It's more within than between hemisphere so very few connections between hemispheres. You also see that those connections are more frequent within modules. The dark gray shows you the borders between modules that are detected in the network. So then we have to normalize because there are more connections within modules, there are more connections within a hemisphere so we can normalize for that but we still find that there was a preference for reducing those fiber tracks. There's a preferential loss of streamlines within thick short distance inter module and inter hemispheric fiber tracks. So if we turn that around the fiber tracks that are not changing during development are the ones that are long distance, the ones that are linking the two hemispheres and the ones that are linking different modules. We didn't test it specifically but presumably the fiber tracks that remain could also be part of the Witch Club or the Formal Networks. That would be something to test in the future. And this is very important because fiber tracks that connect hemispheres and connect far away regions are very important for integrating different kinds of information. They're very important in functional terms. What we also found is that this change, this reduction in streamlines occurs earlier for females than for males. So the pattern is similar but the time course is just earlier in females and one follow-on question is does this link to psychiatric diseases that sometimes occur earlier or more frequently in one gender compared to the other? This is looking at models of global connectivity and we are now using that as a starting point for computer models to change connectivity and look at effects in the system. But we can also look at the micro level, at the formation of connections and layers at the local level. So what you see here is the formation of cortical layers. So you have the subventricular zone, you have neurons that are growing, some neurons are just dividing and others start to migrate. I'm going to the top here and you can see the formation of different cortical layers. We're using that as a starting point to understand what parameters do you have to change to get differences in cortical thickness and differences in the thickness of individual layers because we know from histological studies that the layer architecture has changed to different brain diseases. We're currently working with a certain open lab at Intel in order to scale up those simulations. What you see here might only be a couple of thousands of neurons, maybe 10,000 or so, but the idea is to scale that up to larger parts of cortical tissue and what we can then do is while they're looking at the formation of layers, looking at the formation of jewelry and sulky and the formation of fiber tracks between brain regions. This is the next step by scaling up those simulations. There's more information about our partners on the website here, but we're at the early stages of scaling up those simulations. The next step in treatment is identifying treatment targets and those are not necessarily the ones that show change connectivity. So if you see a change in connectivity, it might be an underlying cause of a disease, but it might also be a consequence of the disease. You might have a brain disease that is happening and because of that, other parts of the system are changing and some of those changes might even be beneficial. They might be compensating for lost function or they might reduce seizure likelihood. So we have to find out which parts of the brain to target and for this, we developed one technique which is looking at higher resolution of targets. This is one example looking at MEA recordings. So you basically have with sector tissue, you have the surface of the tissue, you have some kind of MEA on top of the tissue and you have some underlying sources. So this would be layer one, layer two, three and then the lower cortical layers. So you have some underlying sources and those electrodes are then measuring different time courses. So the idea is that you have the recording phase. So those are the sources and you have local field potentials, you have potential changes at the electrodes, but then based on the measurements you're getting from the MEA, you can do the reverse. You can basically play back the recordings into the tissue. So we call that near field a holography where you basically have the sources of the signal being the MEA electrodes and then the signal is traveling back and the focal points are then the reconstructed sources where the signal is coming from. So this was published last year, but the student who was working on this, Henry Kelston, has continued to apply this to EEG. So while then looking at resector tissue, looking at EEG recordings and the principle is the same, you basically have some sources in the brain that show activity and this activity is then measured in EEG electrodes and by playing the signal back into the tissue in a computer model, we can try to find out where the signal is coming from. Of course, the benefit is that we have a very high temporal resolution. It could have 20 kHz recordings, for example, but at the same time we get a very good spatial resolution including deep brain structures. Just to show you a video here. So this is based on this measure in EEG. That's EEG today. So you have electrodes on the skull and some measurements, but then using this new technology which is currently being patented, we can go beyond those measurements to higher resolution. So 128 electrodes, 256 Hz. That's a measurement here. So you can basically see energy flows in the system, not just the surface of the brain but also in deeper structures. So you have an image presentation as a consequence of the image presentation. You actually see a response in the system. There's spontaneous activity. You have image presentation. You have a gap between image and the response in the system. So you can see after 170 milliseconds you see a response in this recording here. So it's really early stages, but what this potentially gives us is a better refinement of what parts of the brain need to be targeted by invasive or non-invasive brain stimulation. And as part of that, we also need to systematically test different treatment options. So we are developing a workflow model where we can basically test many different combinations. So targeting five or ten different nodes. And as you know, from combinatorial explosions, there are many different possibilities how to target five nodes out of a thousand or out of 50,000. So we're using some optimization routines to find out what are the best combinations for treatment. And talking about treatment, finally, we need to know the effects and side effects. So we already heard about effects of treatment, but it's really important to know the side effects as well. If you only look at effects, a computer model might suggest, well, in order to stop epilepsy for this particular patient, let's take out the language areas because the computer model tells you that this is the best way to stop seizures. But of course, you wouldn't want to take out the language areas. So you need to have some idea what the side effects of treatment are. I just want to mention some applications of looking at effects and side effects. One is a project looking at an optogenetic stimulation in epilepsy patients. There's an implantable device that's implanted in the brain in the cortex here. You have ongoing measurement through normal electrophysiology electrodes. And then you have a decision to make whether to give a feedback signal, whether to stimulate part of the brain or not. And the stimulation is using a light. So you first have a virus that is transfecting some of the new ones. And later on the implant of the device. And when the light is shining, you can basically change the activity levels in the system. And we are developing computer models to decide what kind of stimulation, what kind of feedback is needed. And especially deciding when to give feedback at what stage in the system. And there is a postdoc position available if you're interested. In order to do that, we can use models that were presented earlier by you, so Wilson Cowan model, for example. But we can also think about more detailed models in the future. And one tool will be developed as the vertex simulator as a mudlap toolbox that you can download from the website here. Which basically means that you can simulate a column 100,000 neurons on a normal computer. You don't need a big cluster. And the reason why this works is that we're using a simplified architecture rather than having 100 or 200 compartments for each neuron type. We use a simplified version of five to 10 compartments. But we still get the same local field potential in the system. So testing this model, we compare the simulation with recordings in the macaque monkey. On top, you see the activity that is measured in resected tissue. You have a multi-electrode way. You have the upper cortical layers, lower cortical layers, and gamma activity. So you see higher gamma activity in the upper layers. And you also see a phase inversion between the top electrode and the next electrode here. And we see the same in our computer model, so high gamma activity in the upper layers and the phase inversion between those electrodes here. So it's based on what we know about anatomy in the system. And it can reproduce the local field potential here. And that can be a starting point to test what are the effects of stimulation at the local level. And this you already heard from Peter and Eudrun, so basically using computer simulations to predict surgery, outcome, and alternative ways. So in some way, I think it's really important to have all three components to have some idea what kind of subtype of the disease is present to identify potential treatment targets and to model the effects and side effects of treatments in a computer before treating patients. And this slide was here several times, but I would just like to point out that we have an open faculty position in UCAS, if you're interested, just have a chat afterwards there at the moment 30 faculty members working in that area. And thanks for your attention. Suppose that you want to simulate essentially the connectome to study the effects of a disease, or do you think that you would simulate the individual nodes as oscillators? What kind of oscillator model would you use, like Kuramoto or something more complex? Yes. So basically it depends on what you're interested in. So I think for a lot of the applications we were talking about, we need very simple models. Wilson Cowan, the one that Eudrun was presenting earlier. It's only if you're interested in information processing. So if you say that, well, you have different information processing, different types of cognition for different patients, you need to be a lot more detailed. So in that case, you might need a model that has spiking neurons that is using information theory and so on. In our case, we're interested in more general changes for diseases. So for schizophrenia, it's known that different brain rhythms are different for schizophrenia patients in some parts of the brain. So we're just looking at oscillations rather than looking at detailed information processing. So we're not interested in tasks, for example, but just the general pattern you see at a resting state in patients. Thank you. Yes, yes. Discommunicate RSM-DGs for our life. So we know that many studies have showed the disruption of lowering connections in RSM-DGs. Why didn't you use lowering connections through these mini-analyses? Well, it's basically not something we found for the functional connectivity in our data set. So I think the reason is that we're not, in that particular application, we didn't look at the difference between a healthy controls and Alzheimer. So that's, as you say, it's like long distance. We looked at the difference between Alzheimer and Lewy body dementia. So if both Alzheimer and Lewy body dementia show changes in long-distance connectivity, it doesn't show up because we compare between the two and not with healthy controls. So I think that's the reason why it doesn't show up. Yes, yes, yes, exactly, exactly, yes. Okay, okay, thanks.