 Okay, so now we move to next speaker and we go back to academia now. So, Raffaello, maybe. Can you hear me? Yeah, yeah, yeah. Maybe, okay, so maybe you can share and then I'll make it short. I will start trying to share this slide. So hopefully you see the title slide now. You see, yeah, we can see. Okay, so next speaker is Professor Raffaello Bottestio from the University of Trento in Italy. So just a few words about Raffaello. So Raffaello got his PhD in 2010, right Raffaello, with under supervision of Cristiano Micheletti here in Sissa. Then Raffaello moved to Mines at the Max Planck Institute for Polymer Research in the group of Kurt Kramer, where basically he started his work on, let's say, course-graining, journey course-graining and on molecules and bio-molecules. And then Raffaello got a position in Trento and also an ERC a few years ago, and so with which he started his own research group there, and now he's an associate professor at the University of Trento and today we'll tell us about communication pathways in an IGG for antibody on a multi-scale study. So Raffaello, please, the floor is yours. Thank you very much, Angelo, also for anticipating the customary presentation slide that it's not something I usually do, but of course in this alumni meeting it's a pleasure for me to stress the fact that my career path took me to Trento through Trieste. As you might know, people in Italy think that Trento and Trieste are basically five centimeters apart. This is definitely not the case. Unfortunately, we have mountains but not the sea, and after the University of Rome, I went to Sissa for the PhD, and I think it is worth mentioning in which conditions I was when I started the PhD, which is this, and this is the condition in which Christian decided to accept me as a PhD student. So I cannot but thank him for making this leap of faith, and I agree with Andrea that a PhD in Sissa certainly does you good. And I'm very happy to be here today telling you something about the work that we have been doing in my group here in Trento. As you can see here the happy faces of the people who have really carried out the work that I'm going to tell you something about. So I think that during the last, the past few years you have seen this image several times. Of course this is the COVID virus. I'm very happy to let you know that I will not be talking about the COVID virus today. Actually, I will be focusing on something, however, related somehow to that, which is antibodies. And in particular, I will focus on one antibody which is Pembrolizum, which is a modified antibody. An antibody which of course, which occurs naturally in cells of several organisms, however it has been engineered to act as a drug for pharmaceutical needs. The reason why it has been engineered is that this antibody in the white type can undergo a process in which the two arms, the two arms detach. You should see if you see my quantum you see here to yellow lines. So these two yellow lines are not there in the in the nature of antibody because they represent system bridges called bridges which have not, which are not present in the in the nature of antibody. And because of that, the these antibody can can separate between the two arms and undergoing exchange with other antibodies for pharmaceutical reasons people had to engineer it so that it stays there so that the two arms remain remain there. And it is customarily used typically for several kinds of treatments, most of which involve cancer somehow. So this is a fairly huge molecule, it is a modern 1300 residue large protein it is a multi multi chain protein. And the first step of the function of the molecule you had the binding with an antigen in particular one of the main antigens. And this protein is this PD one protein which is part of a larger complex, and this interaction happens here, which is the region where the, where the antigen binds here you have the binding site the variable in the variable region of this chain of this of this large molecular complex. Now what we did was to investigate the the inner life if you want for this molecule by means of molecular dynamics in relation by means of all out on molecular dynamics in relation so we want to understand something about the, the internal dynamics of this thing, because I will make a step back, you see that this, let's say three blocks three large macro blocks are connected together by by him by hinge, which is composed by two relatively abstracted segments of the other proteins, which allows them a large amount of flexibility. So the question was, from our side, what kind of flexibility is this, is there some relationship between the, the dynamics internal to the various domains and the way the system arranges itself in the, at the large is there really a relationship between the binding state of the of the antibody and it's large scale conformational arrangement. In principle you have three balloons which are attached together by threads and you could expect that no correlation exists. Of course the hinges ready be short for this antibody you might expect something so we wanted to investigate this something. And we did that, performing a relatively large amount of molecular simulations so we have for each state apple and hollow so we have the unbound antibody and the bound antibody bound to the PD one protein. And for each of these two states we performed, we performed for molecular dynamic simulations, which are independent, however, starting from the same initial configuration, but of course with randomized velocities, and these four trajectories are 500 nanoseconds long each. And the, the first step in this in the analysis of the of the system was to perform a clusterization of the frames that is, we consider the frames of the trajectory of the four trajectories for each state all together. And we try to to cluster the conformations together so as to identify macro states. The, the clustering to place through hierarchical clustering strategy which is a relatively way of putting together conformations based on their distance which is of course given by the routine square deviation. We employ the threshold that allowed us to identify a decent quotes number of clusters that is a number of clusters that is large enough for the clusters to make sense to distinguish different configurations. But at the same time it's not so large that eventually you start breaking down clusters that are only, only slightly different from each other. So this is a very qualitative way of making this decision and with clustering typically this is the case and I think maybe Giovanni and Alessandro might confirm something about this but eventually we ended up with this partitioning. So as a first thing you see that there is a difference between the apple in the whole estate in the number of classes in the apple state you have six clusters, which have almost the same population if you want to have three classes that are the same size, and other three clusters which are similar in size among them but smaller than the than the first three, and the whole estate on the other hand is populated by four clusters, which roughly share we share roughly equally the size with a fourth cluster which is relatively slightly populated. If you look at the properties of the of these clusters which of course contain together conformations that come from old trajectories at the same time. So for example you see the cluster three here takes frames that come from the four trajectories independently. And those four and five here only include conformations that have been visited in the fourth trajectory and so on and so forth. So if you look to into the classes and try to figure out what happens in there, you see that there is a substantial variability regarding the apple state, you see that conformations cover a large number of radii of duration and dispersion within each cluster, which is given by these bars. It is interesting to note that the initial confirmation from which all trajectory start is about here. If I'm not mistaken it is precisely this within this cluster here. So you have that whatever starting from the same configuration, you have clusters that represent conformations, which are populated by conformations that tend to expand starting from the initial configuration and other classes in which the system tends to, I wouldn't say collapse but certainly shrink in terms of confirmation of the overall arrangement of the of the structure. In the whole of state something similar happens, but not quite as much as in the apple state so you see that the conformational variability of the whole of state is way smaller is, I would say half the size of what you have in the apple state. So in a sense, this would suggest that the the binding with the with the antigen favors and more rigid a stiffer conformation for the for the antibody and you see that in the number of clusters, as well as in the variability the conformation of variability that you have in each cluster. If you look at the binding site, which is of course an important thing to to to understand to figure out whether some relationship exists between these between the binding and the overall arrangement of the antibody. The in the apple state the binding site has a certain confirmation of variability so you have the routine square deviation from the conformation of the binding site in the initial configuration, which cluster by cluster covers a broad range of conformations and in particular, the, let's say the maximum these distributions are not necessarily in the same place. This is not the case for the whole of state in which you have a slightly smaller range of routine square deviations. But most importantly, you have that the bulk of these distributions is definitely more concentrated in the same region for the four clusters. region which is populated by the apostate as well. So this suggested us the possibility that the, that the apple state and the whole of state shares something from the point of view of the binding site, as well as from the point of view of the overall arrangement. And indeed, if you see, for example, what happens to cluster zero H this is the routine square deviation of the binding site in the in the first class of the zero cluster for the whole of state, which corresponds to this conformation here. Now there is a conformation among the ones in which you do not have the binding site the, sorry, the one, the antigen, which is structurally very similar to this one, which is, sorry, this one. So the one a conformation, even though it might not look. So, just looking at the conformation, it turns out, quantitatively, that this structure is the closest among all the others to the structure at the global level that you have in the zero cluster of the, the whole of state. In the, in the whole of state, you have a distribution of configuration in the active site in the binding site, which is, sorry, in the opposite, which is this, so cluster one a, you have this population that the black line for the, for the binding site, is different from what you have in the binding state. However, the moment you remove from this conformation, the binding site. Sorry, again, the, the, the antigen, the PD one, you have that even though the conformation at the global level remains in proximity of where it started that is this cluster in a simulation of in a few simulations of 100 nanoseconds of length, the binding site immediately relaxes the conformation, which is closer to what you have in the, in the corresponding conformation in the upper state. So there is something that tells you that if the antibody is in this conformation in absence of the, of the antigens, the binding site would like to be in a certain conformation which is not the one in which it is when, it is in interaction with the, with the antigens. So this suggests that there might be some relationship between the conformation of the system at the, at the large level at the level of the overall arrangement of the antibody and what happens in the binding site to figure something out about this in a more detailed and more quantitative manner, we started making use of, of an analysis which relies on mutual information. Mutual information is a quantity that you can compute that allows you to, to say something about the correlation between two to residues in this case, beyond what you would get in a regular PCA regular principal component analysis because this, this goes beyond linear correlations. So you, you have the possibility of computing joint probability distributions that, that are computed on the displacement of the residue, for example, the I residue from its average position. So this is the probability of having residue I displaced by a quantity Xi distance Xi from its equilibrium position from its reference position and this is, of course, the equivalent for residue J. If you compute the mutual information contained in this probability distribution by comparing it by means of what is effectively a quibbler divergence. You compare it with the, with the product of the marginals of these distributions. So in a nutshell, if you have no correlation, the joint probability distribution is equal to the product of the marginals, because it is exactly the product of the marginals. If there is correlation, that the product product of the marginals does not entail, then this quantity is larger than zero. And the nice thing is that you can employ this information is mutual information to construct a network to build up a network of interactions within the of correlations if you want within the, within the system that you can employ to highlight pathways of communication within the system. I put a question mark in the alosteri here because of course this is a is a dangerous word to employ. So I wouldn't say that these are necessarily alosteri pathways, but certainly there is some interaction some communication that takes place within the system at the equilibrium level of course we have analyzed this in the equilibrium quantities, but the result that comes out certainly suggests that there is, there is communication that takes place from the binding site of the antibody at least the one that is, that is more free to move. And the rest of the antibody, why did I mention the one that is more free to move. I did that because this antibody, because the initial configuration that we employed in our simulation is of course the crystallographic one, in which the antibody is arranged in a tilted conformation that allows for a substantial interaction between the, the fab one and the FC domain. So in this respect we consider these binding sites to be to be the one that has larger accessibility to to the, to and indeed we see that it is from there that you have the largest amount of pathways not necessarily all the pathways for example here you see that pathways start from these other variable region as well. But certainly two things are striking in this in these images, one is the fact that there are substantial communication powerful so there is, there are channels that are definitely more populated than other. But bonds that you can identify in the system. And the other thing is that several pathways go through the, the hinge some pathways do not go through the hinge so for example they take a shortcut through the interactions that take place directly for example from between one fab and the other but most of the interactions, especially in the, in the bounce date, take place through the, through the hinge. This might sound slightly of using the sense that the hinge is the part that physically connects the the fab chains that the fab domains to the FC. However, you see that alternatives are possible so it is certainly something to highlight something that comes to our attention that there are these communication pathways. And indeed we focused a bit on the role that the hinge plays, and you see that when the hinge is bound to when the antibody is bound to the, to the antigen to the PD one, there is a reduced flexibility in the hinge so the hinge becomes more rigid. In a sense, this tells you that something changes in the in in the hinge which is a consequence of the of the higher stiffness of the more of the larger constraints or restraints that you have in the binding site because of the presence of the antibody. The hinge on top of that is characterized by the presence of high centrality residence centrality is a measure that in rough theory tells you how many pathways pass through the network pass through a certain set of, of notes, and in the hinge you have a large amount of these nodes. So the, these pieces of information suggests that the, the hinge provides a response to the binding site might have consequences in the behavior of the of the rest of the system so this, the modifications that they place in the hinge reflect on the arrangement of the system as a whole of the protein as a whole. And, and of course, the moment in which you have something something taking place in such a relevant part of the antibody as the binding site. These modifications are communicated through the rest of the, the antibody, mainly importantly passing through the hinge. Additionally, we noticed that if you perform a calculation and approximate calculation of course, but the binding free energy between the, the binding site and the PD one, depending on the cluster, you see that the strongest binding that you observe is the one that is correlated that takes place in the cluster which is the most correlated so if you consider a measure of how much correlation is there. Among the, within the cluster, sorry, within the domains of the, of the antibody you see that when the, when, when the binding is present when the, when the PD one is in, is present and interacts with the, with the antibody, the strongest binding is the one that takes place in the in the cluster where within each domain, you have the largest amount of correlation, which of course means also the largest amount of rigidity, if you want within the, within the domain. The idea, the suggestion is that the presence of the, of the antigen of the, of the PD one in this case modulates in a substantial manner the, not only the binding site, which is something that one can consider relatively obvious and natural, but also the remainder of the, of the antibody, something that is interesting to notice is the fact that if you compare what happens in terms of correlation at the global level and at the level of domains in the, in the system as by the, by a measure, this correlation coefficient, which is a measure of the overall mutual information that you have within a certain, a certain set of residues, you see that if you consider this correlation that takes place within the domains, you have that this substantially doesn't doesn't change much so the, the presence of the, of the binding site for the different class that classes doesn't change substantially. I might have said the opposite before so I apologize for that. You have that the inter value which is the, the, the value of the correlation among different domains is the is modulated substantially by the price by the different classes, and you have that the strongest, the stronger the binding, as it is the case here for the class zero, the, the stronger the correlation among domains. And these also correlates with other, with other other observables that you have, for example, the root and square fluctuations that take place in the, in the antigen itself is, is the lowest when the, when the binding is the strongest again, this, let's say, correlates with what you would expect for a strong binding, and you have large, large surfaces in the interaction with the PD one, and so on and so forth. So this, this is a first indication this is of course a first work that we have done that we are doing on this on this antibody. And there is much more to do but this first analysis suggests that antibodies in particular this antibody which has a relatively short hinge are not just, you know, three large rigid blocks connected by relatively loose threats. Actually, there is a, there is a substantial amount of correlation within the entire, within the entire system and this correlation is modulated by the presence of, of the antigen. And this can can have also an important important consequences in the, in the design of antibodies so the moment you understand, this is the long term perspective, the moment you understand what kind of consequences from the point of view of the overall arrangement, the binding with the binding rate with the antigen has for the, for the whole antibody you might even conceive the the deploy of modifications in the, in the system in the, in the whole antibody that for example increases flexibility reduces its flexibility favors or disfavors the, the, how prone the the other binding site is to bind with with the antigen, because we have seen that there are these communication pathways that if confirmed if they're, and if we learn to play with them can be exploited for for the design of antibodies with specific, specific features from the So this, this work relied heavily I would say on all auto molecular dynamics simulations and these were tough enough for a system that is composed by not even for 1400 residues and this is, let's say this is not a small protein of course, but it is not even a particularly large protein even larger proteins so for example the, the infamous spike protein of the of the coronavirus is roughly the size of the antibody of this antibody however you have complications due to the, to the fact that there are the glycans to be considered in this case of course I didn't mention the fact that this antibody also happens to be covered in coated in sugar if you want to be covered in, in glycans we performed our simulations without the glycans we performed simulations with the glycans and these are the ones that we are investigating at the moment. This protein also have has the, the issue that it is embedded in a membrane so you have to simulate the membrane and so on and so forth. Then you can go to even larger systems that are entire viruses and of course performing gold auto molecular dynamics simulations for entire viruses is something particularly obnoxious to say the least. In the worst case scenario you simply cannot perform the simulation that the systems are too large, the processes that you have to, in that you want to investigate take simply too much time for a lot of simulations to be performed. In case you can, they still are extraordinarily expensive from the economic sort of from the monetary if you want as well as from the time point of view. You have to invest a huge amount of computational resources which implies long times and, and a lot of money. And sometimes it might be even overshooting in the sense that you, you need to perform a lot of molecular dynamic simulations to figure out something that eventually can be and should be rationalized at a much coarser level. At the level of you have this domain that moves like this and this correlates with this other domain and so on and so forth. So in order to understand how these large systems work from the biological point of view. There are several questions that can only be answered at the level but there is a huge amount of questions that can and should be answered in terms of. If you pass me the term course green terms that is at the level of detail at the level of resolution, which is much coarser than the optimistic one. Of course, having mentioned, coarsening coarsening and course greening this much, it is natural to ask oneself whether cosmic models can be a solution for these kind of issues. That is, can we perform simulations directly at the cosmic level and figure out something about the system. Well, of course we can, and we do that a lot because course we models provide a substantial advantage in understanding and understanding these complicated systems. First of all, we have the simplicity and the practicality of having fewer degrees of freedom, which means shorter range interactions, effective interactions which are smoother larger time steps, fewer forces forces to compute, you can run larger simulation simulations larger systems for longer times, and something that cannot be cannot be celebrated too little or too much is the fact that you have a huge amount of different course we models available. And this is important because the moment you course green. Every system is very specific, and having a huge amount of different course we models of shelf that you can employ allows you to select the particular course we model for that particular problem you have attend. And this is something very important because there is no one size fits all. However, of course, there is a drawback, you are losing resolution. And because of that, you can only observe those processes that you can embed in an effective representation of your system, both from the point of view of freedom as well as regarding the interactions. You have no chemistry dependence, unless at the very, again, course model so I'm a very effective model. And if you have processes that start at a certain high detailed level like all out on, and they reverberate up to the, to the large scale of the system, you simply cannot observe these processes, you cannot reproduce these processes in in a course we're So course we need represents as everything it has its pros and its cons, but it also provides you with the, with a twofold way to be employed. On the one hand, you can perform course raining in the manner that I mentioned that is to construct a model, which is capable of reproducing of giving rise to some kind of behavior that we're interested in. So in this case you have your course grain a system to get a model which is simple which is cheaper, and you see what happens in that model. So in this sense you replace higher resolution simulations with something that is cheaper and effective. And course greening pure provides you with the tools to perform a lot of molecular simulation or something equivalent to that. And cause green to figure out something that goes on in the system, because the reason is that when you perform a lot of simulation you have a large amount of information you have in principle all the information you need. You know that the force field is accurate enough and you have no chemistry, no active chemistry going on in your system. However, you have to figure out what goes on in the system and of course in many cases you know already what to expect, you know where to look in your simulation but in sometimes sometimes you have no way of understanding of figuring out a priori. So you should expect worship you should look to understand what your system does. And actually this is the situation in which you would like to be that is the system does stuff that you do not see face first. First value, face value, and you would like to see that emerge and to understand it once you have performed the simulation. This is the situation in which we were when we performed the analysis of the mutual information pathways. We performed the analysis but we didn't know what to expect, and we saw up posteriori, what kind of pathways we had in our system. The question is, is there a way to simplify the amount of information that you get from an allotment like the dynamics simulation is there a way to filter out the information from the noise, starting from an allotment. And understand the system in a way that is intelligible for us for us so you represent the system in an intelligible manner, but at the same time, in a way that is also somehow algorithmic that is that doesn't require us to put the answer in the question. From this point of view, course raining offered us a way of doing things. And the way we implemented this idea is the mix use of the concept of mapping entropy. Now, what is mapping mapping is basically the low resolution representation in terms of which you look at your high resolution simulation. So once a mapping is just a selection of course inside a selection of atoms that we retain while the rest is simply neglected. In doing this, you can calculate what is called the mapping entropy, mapping entropy is a cool buck libeler divergence, again, which is a distance between probabilities distributions and we have two probability distributions which come both from the all out of molecular dynamics This probability distribution is the is essentially Boltzmann's probability distribution that is the one that we sample that we expect to sample in our O'latom and G. And then we have this probability distribution P bar, which is obtained by looking at the O'latom trajectory with course raining glasses. So this M is the problem is the mapping applied to a particular O'latom configuration, and this probability is essentially the probability to sample a course grain configuration in the O'latom set of configurations. We can see that there is course grain involved, but there is no process of modeling involved in this, because we are just looking at the O'latom configurations in course grain terms, and we get this probability distribution of sampling a particular cause grain configuration, which is the one on to which this particular or not configuration maps. We have a normalization factor which is which simply tells you how many configurations you have that map on to this particular course grain configuration. So this is a sort of Boltzmann probability distribution flattened ironed so that you have the same probability the same average probability for all those O'latom configurations that map on to the same course grain configuration. And this allows us to look for the optimal mapping which is the one that minimizes this probability distribution, because once you have flattened the O'latom probability distribution, you can look for the for the selection of sites that makes this distance between probability distributions the smallest, which is to say, the moment I have the moment I look at my system in simplified terms, what is the simplified representation that that that describes my system with the highest fidelity with respect to the high resolution representation. And this allows us to perform a minimization procedure and identify those sites that you definitely have to retain in order to minimize in your presentation in order to minimize the distance between the cause grain representation and the all out on high resolution representation. Here I show you the application of this procedure to a particular case the particular protein which is time up in. It is a relatively small protein which is a toxin binds to an ion channel to a potassium channel and essentially kills you it comes from the venom of a scorpion. What I did was to look for the particular selection of sites of atoms in particular that one has to retain in order to minimize this distance between all out on and course grain perspective on your system. And you see the result for the application of this procedure to time up in with different numbers of sites that you retain. So what do you what do you do here you have first of all a bunch of values of the mapping entropy for random choices of your of your site so so you randomly pick say 124 atoms and you calculate the mapping entropy. This is the bulk of the distribution so this is the range of values that you explore. But then you can optimize you can perform a simulated annealing in which you the objective is to find the particular set of atoms that minimizes the mapping entropy and you get if you do this operation a few times, you get several mappings which are whose values of the mapping entropy are located far away from this bulk and our group together in relatively narrow distributions. You can perform this operation for different numbers of different sites. And what you see is that within each group of optimal mappings for a given number of sites and even across the different groups. These are the two atoms which are retained high probability probability which means essentially how often you find that particular out on in an in an optimal mapping in a mapping that is represented that turns out to be optimal from this point of view. So you come to the bright lines that you have that you have in these plots. And interestingly, there are two residues that contain atoms that are often typically retained that are the blue atoms here, and these two residues are the ones that play a crucial role in the binding of this protein to its natural substrate to the ion channel. Note that the, the ion channel was of course absent in our simulation, our simulations were perfectly standard all auto monochromatic simulations on the protein, and there was no information about the binding site or the, or the, the substrate whatsoever. And yet in this manner this procedure, allow us to identify two residues that are particularly relevant for the for the biological function that the protein exploits in interaction with another one. And I skip the application to other two relevant cases of much larger of larger and much larger proteins that come up in. And the results, the results are consistent. What turns out is that mapping entropy which contains information about the energetics of the system, but it doesn't contain information about the interactions with the substrate is capable of figuring out spots that are particularly relevant for the, for the biological function of these proteins. So this tells you, somehow, if you want to construct a model of your system in which you modulate the distribution of detail on your system in a non uniform manner. You have to retain for sure this view site. So you, you might construct a model in which you have a few residues that are optimistic, and the rest which is course when which residues should I use. Certainly, I have to keep these residues at the allotment level because they are the important ones for the biological function of the product. So the next step, of course, is once I have identified an optimal mapping, I would like to construct a course model onto that. And the, the strategy that we are developing to do that in a, if you allow me, we can dirty manner is to construct a model that starting just from the, from a static structure of the protein, a given force field and the selection we call this canvas, which is the course main unisotropic network model for a variable resolution simulations, and the variable has to do with the fact that you can provide a mapping in which you have varying degrees of detail so that you have certain regional the system that is described at the allotment level, then you have intermediate levels then you have course with the levels. The construction of the model only relies on a particular structure of the protein itself or the system itself and the mapping that you provide. How this work essentially you start from a protein which is represented here with substantial huge usage of fantasy as these sites, and you decide that out of those atoms, you want to retain just two of them, the green ones, and the red ones would be discarded. You want to construct a cosmic model in which those two atoms are there, they represent themselves but somehow they have to account also for the atoms that are not present in the model. How do you build interactions in this case, we do that in a very brute manner that is to associate the atoms that we remove to the atoms that we keep in a sort of Boronoi like strategy that is, if this atom is the closest to this one among the retained atoms then this atom will somehow take care of this. And of course this holds for the other ones. You remove the atoms you are not interested in and then you have to endow these residues with these atoms with the properties that they such that they account for the elimination of the other degrees of freedom. In principle, if you wanted to do that, let's say in a kosher manner, you would have to perform a lot of molecular simulations, and by some kind of course raining procedure bottom up course we need procedure. You would have to parameterize the interactions and the properties for these sites. You do that in a very, as I mentioned, brute and crude manner that is the following we employ a perfectly standard force field with charges with Lena johnson and so forth. And for each of these atoms that are retained and keep track of the of a certain block of atoms that have been removed. The charge is the sum of the charges. The X the Lena johnson epsilon is the geometric average of the excellence of all the atoms. The sigma the Lena johnson sigma is twice the radius of generation of the groups of a group of atoms that the site represents. It's absolutely trivial to do. And, and it allows you for and it allows for an immediate parameterization of these otherwise very complicated, complicated force field because these employees just the same functional forms that you're having a perfectly standard a lot of force field. And it is instantaneous in the sense that you have the static structure, you decide which items to keep your press a button and the script tells you what kind of interactions you have them for. It can be combined with a lot of solvent, no matter what, and, and it is very immediate because you have interactions between water molecules ions, and these sites straightforwardly. You can perform a simulation which you have this modulation in the degree of detail across the system. For example, you can apply that to our pet protein which is kind is, and we perform simulations, these, the ones that I'm presenting now are absolutely fresh results, like I got the plus yesterday. Simulations are not extraordinary long for this for the model. As of now, like for example, six, 16 nanoseconds for the kind is in this. In this model, we have the, the lead and the, which is this and the NMP domain, which are treated at the, the course with level and the rest is optimistic and you see, even though the structure is tilted the orientation is tilted you see what the system looks like, you have this large part which is optimistic, and then you have the course green regions and the optimistic goes into the cosmic region through a region which is an intermediate level of resolution in which you have all the elements of the backbone only, and the rest is not is not present, and then you have only the C alpha atoms, and you see that there is a substantial correlation at least as far as the routine square fluctuations are concerned between the full atomic simulation which is a long one hundred nanoseconds, and the one that we perform, we perform with the with the canvas model. There is a somehow, there is some deviation between the two, we expect that we want that because, of course, we would like to to sample something without a model that because of time, for example, we cannot sample with an a lot of a model. So here you see that the, the lead that fluctuates much more than it does in the whole atom simulation, the NMP by fluctuates slightly slightly less but again simulation in short, what matters for us as of now is the is the trend is the correlation between the, first of all the stability of the model because again, it's super cheap to construct this model and given the way we construct it. There is absolutely no guarantee that it's going to work and these already points from our perspective in the, in the right direction. There's a huge amount of parameterization for the screens and the, and the bonds in this model that is much less trivial than one could expect so these results are actually comforting for us, and even more comforting is the application to the aforementioned mechanism, which is moded in this manner, you have the olatum region which is essentially just the, just the hinge, then you have a medium range region in which as I mentioned, only the backbone of the, of the residues is the main and then you have the rest in which you keep only the, the, the c alpha atoms. And you see here the results again super fresh results very short simulations for the moment for the 48 nanoseconds, definitely not much for a system that large, but we see satisfactory, if you allow me correlation, certainly not in the amplitudes amplitude for example, you know our 500 nanosecond, sorry 200 nanosecond simulations in this, in this case are much larger than what we see in our much shorter simulations with the, with the model, but the trend is there, we see that there is a correlation between the RMSFs that we compute and in particular, in particular what matters for us is what happens in the, for the olatumistic residues which are the black dots here, which are sufficiently close to a straight line that doesn't have to look one, because again the overall amplitudes are different, but the, what is important is that there is a correlation between the two and of course this is all work in progress. So a lot has to be improved but we hope by the end of the year to publish some, some update on that on the, on the archive to conclude, we have seen that antibodies at least this one with the one that we have been working with are very, very dynamic people, they do stuff and they do stuff, showing some correlation that takes place within the system. That are, that are not trivial, something that you wouldn't expect to begin with and an important role in the, in the system we've been studying seems to be played by, by the hinge in a less trivial manner that what one would expect. Then we have that course-grainning is cool, not only because it allows you to construct simplified models and to run faster for larger systems and so on and so forth, but it is actually something that can be employed as a strategy to, to filter out the information that you have, like your dynamic simulation, even though you do not know a priori where to look. And this is, from my perspective, the, the most interesting direction which course-grainning can move, that is, not only you understand your system because you have simulated it cheaper, but also understand the system because you have performed a lot of simulation and by course-grainning in the right manner, you filter out something that is important, not for you, but for the system itself, and there is a big difference between the two things. And of course, as always in life, one has to pay attention not to overdo the simplification, but of course this, this is a general issue with, with modeling. And with this, I would like to thank you very much for your attention and if time allows, I'm very happy to take some questions. Thank you. Thank you very much, Rafaello, virtual club for you. No questions? Lunchtime is approaching. Yes, yes. Oh, you can also use the chat if you want, I remind you. Sorry, maybe Rafaello, I have a question, my set. So, can you, I mean, with this course-grainning method, can you, you know, I mean, can you somehow simulate also the dynamics of this and how realistic is it? And with the canvas model? Yes. Yeah, actually, so we perform standard, if you want molecular dynamic simulation with a model. Sorry. So the sampling, the sampling strategy is, is, is regular MD. We also perform also an explicit solvent, which from the point of view of gaining time or saving time certainly doesn't help. And this is the reason why the next natural step for us is to combine this method with a multiple resolution strategy for the solvent in which since in a lot of region for your protein and course in region, you treat the solvent all out on only when you have the lot on part of the protein and course going to rest. And, but the eventually you, you perform, you perform perfectly regular and the, and of course as any course-grain model dynamics from the point of view of time has to be taken with a huge grain of salt. However, if you want to, to see for example, conformational transitions, our hope is to, is to come up with a, with a method that allows you to see otherwise hard to observe conformational transitions in large systems. There will be one of the main applications of this, this approach together with, for example, binding free energies calculations. Okay. I'm going to chat. Other, probably there is a questions, but someone is saying I must be muted. So maybe Patricia, maybe you can write on the, the chat, your question. Yeah, there is someone who wants to ask probably. I heard a very faint voice. I don't, I don't. It's actually better. Yes. Okay, maybe. Okay, so that was my fault. Sorry. I think that that should be it. All right. Thanks for the talk. I was wondering about the last part. So you, you might get or not get the backbone atoms when you do that minimization, right? You're not, you're not biased in favor of the background atoms, right? You mean this or the, right. Yeah, yeah, we have no. So the selection of what atoms to keep at the bottom level what to cause rain and so on and so forth is left completely to the user who have who might have several reasons to make a choice against another. In the model as it is constructed as it is conceived has this region in which you only keep the, the, the, you keep all the backbone atoms, and you effectively integrate out if you want the side chains, but the reason is structural is to, is to keep a smooth transition in the coarser part in which you only keep the, the c alphas and, and the a lot of part, however, there is no reason whatsoever why you should keep particularly the c alphas or all the c alphas. You can also decide to retain say once alpha out of two or only the oxygens or God knows what. So this is, of course, we perform this choice because of simplicity. This is the as a course grading amino acids on to the c alphas is the most standard thing that you can do. But this is not necessarily the optimal choice. And actually, what we have seen before about the optimal mapping tells you that this is not necessarily the optimal choice. Right, and you recalculate the forces between each step. Right. And we recalculate the forces as in any force field, if this is what you mean, but I mean parameters, sorry, not the parameters parameters of the site and once and for all. They're fixed, right. She's a good approach, a huge approximation of course, but the model is course. Right. And, and there are repulsive forces like right, it's not like a normal mode analysis like if you try to simulate, you know, if you either time to time step to normal mode analysis, you end up with overlapping atoms because you don't know. No, you have. So each site that you retain has a charge. If it isn't, if it is no zero, it has a radius. It is an effective sigma and it has a linear jobs parameter. So there is no overlap. There is no, those are by all means, interaction sites with a finite short range distance with a with a volume and so on and so forth. Right. Good. Well, I had a few extra questions about the first part. I don't know if there is time. Yeah, maybe, if you don't mind that there is another question and then we go back to you. Okay. So, so I felt that it's. Clustering analysis section. Okay, of course. So let me, let's see if I can do this in less than 300 years probably not. Okay, I will start talking in the meantime so the, the point is that the, the trajectories that we have the simulations that we have performed. Allow the, the antibody to explore a large variety. Okay, there we go. A large variety of conformations. And they are very diverse. And, and there are a lot of conformations per se. So you have. Okay, so you have for each state, you have two microseconds of simulation. Our, our objective was to, to simplify the overall amount of data that we get from the simulations by clustering conformations, irrespectively of the time sequence. So as to group together conformations that are structurally similar and representative of what we might consider a macro state, a macro state from the conformational point of view, of course the, what you see here are the most representative conformations of the cluster and the, and in each cluster you have fluctuations of the structure about these structures, but you see that there is a substantial difference, substantial microscopic rearrangement among the various states. So within each state, you have, you identify these macro states, and you can compare also between different states, you have the auto state in the whole state and you see that the, the, the course green level from the point of view of the large structure of conformational variability, the antibody arranges itself in a different manner within a state and across different states and then these allows this strategy allows us to perform a simplified comparison between the two. I hope I answered your question. Maybe. At least if you want to ask your question. About this, given that you perform clustering and your interest in correlations between motions. I'm guessing you tried principal component analysis and didn't work. Yeah, so we perform PCA within each cluster because of course, in order for PCA to make sense you have to have a sufficiently decent idea of the reference structure and fluctuations about that. PCA throughout the trajectory wouldn't have made sense for the same reason I was mentioning, but within each cluster you have a group of conformations that move around the fluctuate about a well defined reference structure. We perform PCA in each of them. We calculated the first few normal modes for the, for each, for each cluster. This is cumbersome because you have these many, you have 10 different conformations and for each you have at least three normal modes that you have to look at. Eventually the PCA analysis highlighted some, of course, concerted motions that were consistent with the, with the kind of rigidity that we observed in cluster with the, with the degree of flexibility that the antibody. in its generality, its complexity has, but eventually nothing particularly informative came out of that analysis. It is in the supplemental information of the paper however. Did you calculate RMSD between domains or chains? I'm seeing a lot. So the RMSD values are very high and I'm guessing one like you later said there's a lot of hinge movement and so if you do RMSD on the whole structure and there's a hinge movement you get kind of. Yeah, so you, you have relatively important fluctuations of the RMSD. If you look at the antibody itself at the whole antibody. However, if you look at what happens within each domain, let's say the Fab 1 Fab 2 FC domain, and you look at the antibody within, within, sorry, the RMSD within each domain. You have much smaller fluctuations in particular. If you look at the RMSD of each single domain within each single cluster, you have the RMSD that you would expect in a nicely equilibrated simulation. Yeah, I think we stop here.