 Hello everyone and welcome to the 24th edition of the BioExcel webinar series. My name is Rosen Apostolov and I will be today's host. At today's presentation we will have a look at one of the most popular force fields for course grant simulations, the Martini force field, which will be presented by the lead developer of the force field, Sivartian Marek. Before we start, I have to tell you that this webinar is being recorded and after that we will upload it to the BioExcel's YouTube channel, so that you can look at it at your convenience or share it with colleagues or friends. This webinar series are organized by BioExcel Center of Excellence for Computational Biomolecular Research, and I'd like to give you a short overview of our activities in case you're not familiar with them already. BioExcel has been running for several years as one of the European centers of excellence, and we work with several key applications for molecular dynamic simulations such as Gromax, for integrative modeling Hadoq, and for hybrid QMM modeling CPMD, and we work on improving their performance, efficiency, and scalability. We also work a lot on usability aspects of these applications by devising efficient workflows and coupling them with the necessary data integration systems. We work with very popular platforms such as Galaxy, KNIME, Apache Taverna, Comcess, OpenFacts. BioExcel also provides plenty of services in training and consultancies where we promote best practices for usage of the applications and making the best use of high performance computing and high throughput computing resources. Within the center we run a number of interest groups that some of you might be interested in joining. The interest groups cover several different aspects of the broad field of computational biomolecular research, such as integrative modeling, free energy calculations, workflows. We have an interest group specifically looking at practical applications for industrial users. So if any of those are appealing to you, please visit our website and subscribe to them. We also have support forums and chat channels that we encourage you to use if you have any questions or you would like to get in touch with us. At the end of today's presentation we will have a questions and answers session where you will be able to ask your questions to see where can. For that, during the presentation at any time, feel free to use the Questions tab in the Code Webinar Control Panel. And when we start the actual Q&A session, I will give you the microphone and let you speak directly with Sivar Kian and ask your question. If you don't have a microphone or we have problems with it, I will read the question on your behalf. And of course, you are always welcome to join us for discussion at our forums at ask.bioxcel.eu. And it's my pleasure to present you today's speaker, Sivar Yandmaring from University of Rohingy in the Netherlands. Sivar Kian received his PhD in Molecular Dynamics simulations in 1994 and he did several postdoctoral research works in Germany and Australia. And since 2005 he's a full professor and he's heading the research group in Molecular Dynamics at the University of Rohingy. It's one of the universities with very big contributions to the field of endless simulations. Sivar Yandmaring is also a director of the Berenson Center for Multiscale Modeling in the same universities. His interests are on multiscale modeling and especially looking at the organizational principles of cell membranes. Many of you know him as the lead developer of the very popular Martini force field, about which we will hear in more details today. I would like to say welcome to Sivar Kian and I will now give him the opportunity to present. Okay, good afternoon everybody. I hope everybody can hear me clearly. And well, I assume for some of you it will be a good afternoon, others might be in the morning session or even in the middle of the night. For me it's a three o'clock in the afternoon, sunny weather and I'm happy to talk to you about the Martini force field that we have been developing in our lab over the past years. So to all get on the same footing, I will first talk about a little bit Martini basics, how to prepare actually a perfect Martini. In general, the Martini cocktail combines gin and vermouth at a ratio of two to one and you mix it together with ice cubes and garnish with an olive. Of course, there's a lot of varieties that you can try. There's the dry Martini made with dry gin, the Martini Russell with red vermouth, the vodka Martini where you stir in vodka instead of gin, and the perfect Martini using equal amounts of sweet and dry vermouth. And of course, if you're not so much into alcohol, you can go for a Zen Martini with no gin and no vermouth either. Of course, there's always people skeptic about Martini. They say that just filling a glass with gin, waving it in the general direction of Italy should be enough to cover the Martini. Okay, let's go to the more serious stuff of this webinar. This is the outline of the webinar. I will first talk about the looks of the Martini model. And then I will show you some examples of how we validate our model. I briefly discussed some of the major limitations of the model because it's always good to know where the limits are if you apply a model. But then at the end show you a variety of applications of the Martini model. Okay, so where is this Martini model to be found in the spectrum of multi-scale modeling? Actually, the Martini model is a coarse-grained model which allows you to bridge the old atom to the continuum scale. So if you want to have full atomistic or even quantum level of interactions, you are at this end of the scale where you can maybe simulate for nanoseconds and look at system sizes of tens of nanometers. But eventually if you want to bridge to a scale, say a full cell, of course you have to increase both the length and the time scales of your simulation. And by coarse-graining, by replacing all your atoms or groups of atoms by coarse-grained beads, you can of course achieve this kind of speed-up that allows you, at least with the Martini model, to access really the millisecond time scale and the micrometer length scales. So the idea of the Martini model, the way it has been parameterized, is in a very hierarchical, systematic fashion. Starting from small building blocks, small molecules for which there is a lot of experimental data available and for which you also can run fine-grained or atom simulations. That can be used to construct the so-called building blocks and they can then be tied together to form simple systems. Again, for the simple systems you have access to atomistic simulations and you can use additional experimental data to validate a model. And then eventually you can use this to model the complex systems that you cannot use the all-atom models because simply the systems are too big or the time scales too long. Using back mapping, you still would have access to the fully atomistic details if required. But this is the hierarchical approach that we use to model or to build the Martini model. Okay, let's have a look at the way the model has been constructed. It uses this building block approach. So we realize that there is simple chemical units that you can find irrespective of whether they're part of a lipid or a peptide or an alkane. So in this case we have units of four methylene groups. And those are then represented by a single coarse-grained bead in the Martini model. So the Martini model reduces the complexity of these molecules and considers groups of atoms as building blocks. And from these building blocks, molecules are then being formed like you can build Lego houses from Lego bricks. So on average at the Martini level we have four heavy atoms plus the associated hydrogens that are united into these efficient or effective coarse-grained beads. Of course you need different type of building blocks, different type of these coarse-grained beads, because depending on the chemical nature of the groups that you represent, they can be very apolar. But they can also be very polar or even charged. So the Martini model considers these 18 different building blocks ranging from the very apolar ones, denoted by a prefect of C. And then one to five is a subdivision where level one is the most apolar building block and five is the more or less apolar one. Then there's intermediate polarity building blocks. And then again the polar blocks range from one to five, P5 being the most polar known charged building block in Martini. The intermediate ones, they come in different flavors depending on whether they can be hydrogen bond donor with a sub-script D or acceptor type. So if you have a group that can be hydrogen bond donor, you will have slightly different interactions. And especially enhanced interactions with groups that can have acceptor qualities. The same is true for the charged groups. They also have donor or donor and acceptor capabilities, but on top of that they also carry positive or negative charges. So as taking a lipid as an example, we see how different type of building blocks can be placed on top of this structure to capture the overall polarity of the lipid molecule. So we have the hydrophobic tails represented by the most hydrophobic particles in Martini C1. Then the glycerol linkage of the lipid is modeled by NA particles because they can also accept hydrogen bonds. And then the citriolic head group in this case is represented by two charged particles. So these monobonded interactions between the building blocks, these are of course key in the Martini model, are parameterized based on reproducing experimental thermodynamic data. I will show you in a minute how exactly this is done. In addition to these non-bonded interactions, of course, we have to keep molecules together by a set of bonded interactions, much like you do in atomistic simulations. And these are parameterized mainly based on reference atomistic simulations so that we can match the conformation sample at the coarse grain level to the ones at the Martini level. So having constructed a lipid molecule, of course, then there is a long set of validation experiments that need to be performed to see whether the model actually is accurate and realistic. So what you can do for instance in this case is have these lipid self-assemble and indeed you see they form a nice lipid bilayer that you can also then calculate various properties and then see how well they match experimental data. And I will also show an example of that in a couple of slides. So this is in a nutshell how the Martini model looks like and how is this being set up. So originally this was developed for lipids, but now this model has been extended to encompass actually all major classes of biomolecules like proteins, sugars, nucleotides, but also extended to other non-biomolecules including many types of polymers and other type of nanoparticles like fullerines. Of note is also the water models that are associated to the Martini model. The standard water model is a leather Jones bead that actually represents four real water molecules, similar to the 4 to 1 mapping that is the basis of Martini. And we also have a polarizable version that can also screen the electrostatic interactions if needed. So the key features of this Martini model are summarized here. So there is still chemical specificity, which means that you can really distinguish between different lipid tails, for instance, or different amino acids, different nucleotides. So there is enough chemical specificity that we retain that capability. Of course, the reason for doing Martini simulation is the main reason is that it has enormous speed up compared to all atom simulations about three orders of magnitude. Then this building block approach also ensures that the model is very compatible and versatile. So you can easily combine all these different biomolecules as well as non-biomolecules together in one simulation because all the interactions have been calibrated in the same way. And it's easy to extend this to include novel type of molecules that you want to simulate. So the prioritization, as I already said, is a combination of top-down approach where we use experimental data, in particular thermodynamic data, to calibrate in particular the non-bonded interactions, as well as a bottom-up approach where we rely on atomistic simulations, small-scale simulations, mainly used to calibrate our bonded interactions. You can summarize this as a top-up approach. Okay, so what about this name actually? People wonder, Martini, known as the cocktail. Well, in fact, the Martini force field is named after Saint Martin, which is a patron saint of the city of Groningen, and Groningen is the city that I was born in this city and I'm still living in this city. I'm working at the University of Groningen, so I'm very fond of the city of Groningen. And Saint Martin was a patron saint of this city, so you will see many buildings in the city that are named after Martini. There's our famous Martini Tower, and this is why I thought Martini was a nice name to honor the city of Groningen, but I must admit I also quite like to drink this cocktail. Okay, let's have a further look at how the Martini model interactions are actually working. So the non-bonded interactions are based on the standard interactions that you also find in atomistic simulations, namely Coulomb electrostatic interactions for particles that carry full charge. And you have to note that the Coulomb interactions are screened by an electric constant of 15 if you use the standard Martini water model because that of course cannot perform screening by itself. And then all the other non-bonded interactions are described by Leonard Jones type potential to mimic the dispersion and overlap repulsion forces between atoms or groups of atoms in this case. It's also important to realize that all the potentials are short range, so we use a cutoff of 1.1 nanometer in this model, which means that beats see two to three neighboring beats, and after that the potentials and forces are switched off and vanish at the cutoff. This means also that in terms of Coulombic screening actually at the cutoff the effective dielectric constant is infinite, so we have a kind of effectively distant dependent screening of your Coulomb interactions. All right, so the Leonard Jones interactions actually depend on the hydrophilic nature of the beats. So the type of Leonard Jones interactions that we consider between the beat types come at different levels where level zero is the strongest interaction. So there is a large well Leonard Jones well so particles that interact through this level like to sit next to each other, whereas the weakest interaction is a level eight, where the Leonard Jones attraction is much smaller. And then there is one special type of interaction. So these levels are characterized with well well depth between two and five point six kilojoules per mole, and they all all the beats have the same size. This is for simplicity. The exception is this this temp level that is mainly used to increase the repulsion of charged beats and fully a polar beats. And what is also important is that all the cross interactions between the beats in martini are explicitly parameterized so there's no combination rule, as is usually done in atomistic models. Here we can see the full interaction matrix between all the 18 beat types. So here you have the charged ones, the polar ones, intermediate ones and a polar beats. And the Roman letter denotes the strength of the interaction. So to guide you a little bit through this table. Let's focus first on the before interactions with all the other beats before is actually the beat type that we use to model water. You see that the strongest interaction of water are with the charge beats. And of course, it makes sense that charge beats like to be surrounded by water beats. So this is a strong interaction. And then as you go down the list of polarity, the interaction level with water becomes less and less, and especially the a polar parts only very weakly interact. So and this gives them rise to the segregation of oil and water. So that mimics the hydrophobic driving force of the separation between polar and a polar beats. You can also have a look at the self interaction level, which also gradually decreases if you go from the polar substances to polar beats to the more a polar beats. Usually polar beats, they pack at higher density. They have a stronger self interaction, stronger salvation energies than the a polar beats. So that is all mimics by using these Lena Jones interactions. Of course, there's many more numbers here and they have been mainly optimized by looking at partitioning data. And this is explained at this slide. So what we have done to calibrate all these non bonded interactions is we've looked at how different beat types partition between different type of souls. So in this case, we look at a simulation box full of water beats and octano beats. And we just simply count where polar beats or intermediate beats or a polar beats of all the different flavors are partitioning. And from the ratio of the densities in the two phases, you can then get directly free energy partitioning or transfer free energy. And for this, there's a lot of experimental data that can be used to actually compare and to to fine tune the interaction so that you get correct partitioning of the different beat types between a large variety of different solvents. So and using this approach, you can then construct what we call the martini Bible, which maps beat types, which you see in this left column to certain chemical building blocks. And here you will see then how the partitioning free energy of these building blocks compares to experimental data of these chemical compounds. And if we focus, for instance, on the partitioning free energies between hexadecane and water, you see that we get a very nice match of these beat types between experimental data that are available for these compounds and the calibrated core screen interactions. Okay, of course, there's always things that immediately are clear are not going to be working using this approach. And one thing we realized very early in developing the martini model is that the four to one mapping is actually inadequate to represent rings or benzene ring as shown here on the right has six heavy atoms. And yeah, so you with a four to one mapping, you either have one and a half beats, which is a bit awkward. So you could have one beat, but then of course you totally missed the ring geometry. So what we do there is we map it actually to three beats. And we call them as speed. So there we use a somewhat smaller beat that has a reduced sigma so reduced size, but also reduced interactions with all the other beats. And this has been, again, calibrated based on densities and partitioning free energies for, for instance, benzene and cytohexane. There's also in a later version of martini that was used to model nucleotides. There's a class of even tinier beats that are introduced to the model that allow the correct stacking distances of planar compounds. So these are then called T beats, but these are somewhat specialized beats. So we have the normal beats and beats that follow the normal rules than the small beats, the S beats that are somewhat reduced and then specific beats for specific cases, T beats. Okay, then something about the bondage interactions that we use. Again, there we use the standard type of bonded potentials that you can also find in any atomistic model. So we have bond stretching, bond vibrations, angle vibrations and dihedral to control the overall conformations of the martini molecules. So the way they are being calibrated, as said, here we use mostly reference or atom simulations. So you can run an atomistic model of a small compound that you want to parameterize in martini. And for instance, in this case, you can look at the effective angle distribution that is sampled atomistically and then calibrate your angle potential at the course band level, such that you get distributions that are overlapping as well as possible. And this is usually done in an iterative way. So you start with a certain angle potential and then you modify that in a couple of steps so you get the overlapping distributions as well as possible. Okay, then for proteins, actually the situation is a little bit more complex because here we usually also need additional bonded potentials to constrain the overall secondary structure of proteins. Because one key limitation of martini is of course that directionality of hydrogen bonds is lacking because we average small groups into effective coarse grain beats. So the directional hydrogen bonds are only treated in an isotropic way. And if you for instance look at an alpha helix of a peptide, then there is in reality a lot of directionality, directional hydrogen bonds that keep the secondary structure stable. So in martini, we mimic this by introducing certain angles or dihedral that kind of fix this secondary structure. So the secondary structure in this case is an input into the simulation and cannot change during the simulation. We have a script called martinize that actually can generate these topologies for you. So that's all very easy. You just have to provide an atomistic input structure and then you get out the protein topology and structures at the martini level. To construct larger proteins, actually there is something else that we usually need to stabilize the global fold of a protein. And this is done using an elastic network approach. One of these approaches is termed elnadin, which basically defines additional harmonic potentials between all C alpha beats within a certain cutoff. So this is illustrated in this little rocking image where you can see additional bonds being added between coarse grain beats to stabilize the overall fold. Of course you're free to cut away bonds again to allow certain domain motions in your protein. But this is the way that martini proteins are kept stable and close to their native or whatever other stage you're interested in. Again, this is needed because the directionality of hydrogen bonds is missing in martini. We use a similar approach in fact for stabilizing double stranded DNA or RNA. Okay, a few words on why martini is actually so fast. As I said, three orders of magnitude speed up typically. Of course, one reason is that there is less particles, so there is fewer interactions to compute. Then we use only short range potentials, which also makes it much faster. Then there is less friction and this also leads to faster sampling. So you miss a domestic degrees of freedom, so the energy surface is much smoother. So this allows you to do faster sampling of your potential energy surface. Then on top of this you can also use time steps that are typically an order of magnitude larger than what is commonly used in atom models. One reason actually for being able to use larger time steps is that a very accurate sampling is less critical. And what I mean by that is illustrated in the energy landscape below. You also suppose in reality the energy landscape looks something like this. Then if you have a very accurate or atom model, it should follow reality quite closely. And then of course, you also would like to sample this landscape as closely as possible because all the details are actually realistic. At the course grade level actually you simplify this overall energy landscape. These little bumps or this friction from individual atomistic degrees of freedom are integrated out, so the energy landscape is already much smoother to begin with. And then the sampling can be done with larger time steps because you want to capture the overall shape of the energy landscape and not so much the very tiny nitty details. Okay, an overview of the Martini versions that have been developed over the years. The first version was actually an internal version developed around 2000 starting with only four particle types and four interaction levels originally for limits only. Then the first public release was Martini 1.4, which was an improved version of this initial one still only for limits. Then in 2007 the larger number of particles was introduced. So this is the current standard Martini model that I have just been talking about featuring 18 particle types, 10 interaction levels originally also still for limits only. But then this was soon after extended to proteins in the Martini 2.1 version. Again, redefining some of the basic interactions. Then in 2013 actually there was a Martini 2.2 version, which was again an improved version of Martini 2.1, particularly for proteins and also a 2.2P version that is a polarized version of Martini 2.2 where some of the side chains also now feature. Internal charge leads that give it some polarization effects. Then in 2015 there's actually also implicit solvent version introduced. So far this works nicely for lipids, but not yet for proteins. And then we're all eagerly waiting for the 2018 release of Martini 3. Which is about to be released, expected release date yesterday. It should have been released already, but as these things go they always require more testing, last-minute testing. And we're so we're hopefully releasing that very soon. So it will have even more particle types, more possibilities, more particle types. Also more interaction levels, some more fine-tuning has been done. So some more interaction levels have been introduced. But the major thing has been a recalibration of these normal, small and tiny particle interactions. So they're much better calibrated in this new force field. One of the major advantages is also that the stickiness of Martini proteins. We and others have noticed that proteins have a too high tendency to stick together. And Martini 2 versions has now been rejuiced and largely resolved. So the little movie here shows you actually what we now can do with Martini 3. For instance is to mix different resolutions. So we have here a simulation of liquid Dodecane at three levels of resolutions. Either Dodecane is represented by three normal beats in orange, by four S beats in red, or six T beats in magenta. And as you saw, the whole thing nicely remains completely mixed. Okay, that was a quick glimpse of Martini 3. Something else that is interesting is a whole range of high throughput tools that are also available, like the INSANE protocol where you can quickly set up complex membrane systems. A similar protocol, but now within the Charm GUI framework called Martini Maker. There's the backward scripts that you can use to transform from coarse-grained or supra coarse-grained to coarse-grained and fully atomistic models in a very efficient way. There is the DAF protocol that can be used to sample protein-protein interactions in a high throughput fashion. And there's also an automatic topology builder that has been developed by Burrow and Cramer that can be used for a quick initial guess of making topologies also in a high throughput version. So these are all very useful tools and more and more tools are being developed, not only by us, but also by many other Martini users around the world. Okay, then now I want to quickly go over a few examples of how we validate the models that we make. So again, this is comparing Martini to experimental data in case of properties of lipid membranes. So we see, for instance, that structural properties like the area of the lipid for different lipid types are very well reproduced, but also elastic properties like bending rigidity or area compressibility are in the right order of magnitude. Thermodynamic properties like phase transition temperatures or the line tension, the tension you need to build a membrane edge, for instance, are all semi-quantitatively reproduced. And the same is true for dynamical quantities, although usually we're happy there with order of magnitude correspondence. Another important type of characterization is looking at lipid phase behavior. So what we can do is actually we mix, for instance, saturated and unsaturated lipids that experimentally are known to separate into liquid ordered and liquid disordered phases. And we can mimic that in our Martini simulations. So we saw in the movie that starting from a completely random mix of these two types of lipids together with cholesterol, they separate into a liquid order and a liquid disorder domain in perfect agreement with the experimental measurements. Another example of validation is in this case, comparing to all atom data performed in the lab of Tilemon for the partitioning of amino acid side-chain analogs across lipid bilayers. So here you see the free energy profile of dragging a leucine from the water phase across the bilayer interface into the middle of the core of the bilayer. And you see in this case because leucine is a hydrophobic side-chain. It likes to sit in the membrane, but there's a large or small, well, a reasonably small barrier for actually entering it. And you see that the coarse grain model captures this behavior quite accurately. The same you can do for other amino acids. And of course we looked at all 20 possible amino acids. So here's the example of serine, which is a slightly hydrophilic amino acid. And you see, indeed, there's an increase by about 20 kilojoules per mole by partitioning from the water phase into the core of the bilayer. For glutamate, charge 1, the increase is even larger. And there you see that we're actually missing some of this interaction at the coarse grain level. But provided that the charged residue you will actually never find inside the bilayer. So the first part is actually the more relevant part. And this is quite well captured at the martini level. And here's another example where you see how for aromatic residue also globally we capture the same profile as seen atomistically, although there's certainly still some room for improvement in this particular case. Another validation example, again, looking at coarse grain simulations and comparing this to atomistic data. So what we see here is actually a phospholipase binding to a lipid membrane interface only shown by the head groups. And here cation pi interactions are actually important. So the interactions between the co-line head groups of the lipid bilayer and these ring, the aromatic side chains is important. And this is one of the features of the new martini 3 model. So this was a martini 3 simulation. And you see if you look at the overall binding profile at the depth of the anchoring of the residues and comparing all atom data to the martini 3, we see that we can capture the way the protein embeds itself inside the membrane very accurately. Okay, what else do we have? This is an example where we validated how well martini could capture structural changes in protein. So remember that the secondary structure in principle is fixed. But in this case we have a kind of sensitive channel that is consisting of separate trans-membrane helices that can move independently from each other. So each of these five, it's a pentameric channel. So each of these five helices has its own elastic network, but they can move independently from each other. So what is known experimentally is that if you put a membrane under tension, this channel actually starts opening. And this we can mimic at the martini level by putting tension on the membrane. And then the movie shows indeed that the channel opens under this tension and you see blue beads, coarse-grained water beads now being able to escape from one side of the membrane through the other. So this is another example of validation where we know that the protein should open under tension and we can mimic this process quite realistically. Okay, here we have another example of validation. In this case we're looking at cytochrome PC1, which is a respiratory chain complex that you find in mitochondrial membranes. So we set this complex up inside a membrane containing special lipids, cardiolipins that you find in mitochondrial membranes. And here you see a top view of this whole complex. It's a dimer complex. And here we're going to look at how these cardiolipins diffuse in the membrane and are able to bind to this complex, but also in particular that they can bind to sites that are identified as binding sites in the crystal structure. So these cavities, they apparently like to bind cardiolipin. And in the simulation we can actually see this happening. We see a gatekeeper that moves out of the way and then eventually cardiolipin molecules are able to spontaneously bind to both of these cavities in perfect correspondence to the experimental crystal structure. So this is again an example of validation of our martini mole. Okay, so now I want to spend a couple of slides on some limitations of our model. And one of the limitations has to do with the semi-quantitative nature of the model, which is also called the fuzziness of the model. So suppose that you want to model a DMPC lipid, which is a typical lipid which has 14 carbons in its tail, then you immediately have a problem. Of course, you can use this 12-carbon tail, which has three beats in its tail, or you could resort to one tail bit more, but then actually you're representing a DPPC lipid. So what do you do in this case? Well, you can look at some properties. So you can say, okay, let's have a look at the bilayer thickness. Experimentally for DMPC it's 4.3. It's actually quite close to the three-beat martini model. But then if you look at the melting temperature, the gel-2-liquid melting temperature, then actually the four-beat model is closer. So then again, what do you do? So the solution is actually that there is no escape here. So you have to depending on what you want. So if you're after structural data, you probably go for the three-beat model. And if you want to have the thermodynamic data more correct, you might use the four-beat model in this particular case. But you can use it to your advantage. So you can say that you wrote in the paper that your DPPC bilayer melts at 295 Kelvin. And then the referee says, oh, that's 20 Kelvin too low. That's not good. And then you say, oh, sorry, I actually meant I modeled the MPC. And then the referee immediately would say, oh, but then it's fine because then it's bottomed and I'm happy to accept your publication. OK, another limitation that I already discussed a little bit is that we have the directionality of hydrogen bolts missing and therefore secondary structure of proteins and nucleotides is being fixed. We do this using elastic networks. And therefore, one major limitation is that protein folding, for instance, cannot be simulated. There is a potential solution that has been recently introduced by POMA, CHIP, TAC, and THEORIDATE. That is using a gold martini type of network. So instead of a fully elastic network, they use gold type potentials that are also allowed to dissociate and therefore you can mimic the actually folding of proteins using this type of elastic networks. So that's a promising way to maybe get more structural flexibility into the martini model. Another limitation is that the friction is missing or that is an advantage because that allows you to do much faster sampling. But that means also that the timescale should be interpreted with care because all this friction of all the small atomic details is missing so your kinetics will be artificially enhanced. Usually we find a speed of about 2 to 8 or 10 depending on the system details. But also in complex processes, of course, the kinetics might be different more because there depends majorly on the barriers between states. But this is also true for atomistic models. So in general you have to be careful by interpreting the kinetics. So timescale should be interpreted with care. A solution is of course to apply friction to the equation of motion. You can do this differently for different degrees of freedom. Of course you would then also lose a lot of the speed of the model so in principle this is not recommended, I think. But you can potentially do that. Another limitation that is inherent of any coarse-grained model is the fact that you're losing degrees of freedom and that you're replacing some entropy by enthalpy terms. So you can see this for instance here if we look at the conformational entropy of a lipid molecule over time, a lipid in a by layer you can estimate the conformational entropy through this Schlitter equation. So in a fully atomistic model it builds up to this level and if you do it at the coarse-grained level you see the entropy is much lower precisely because you're missing some of the degrees of freedom. Of course if you compare to a fully atomistic simulation that has been mapped afterwards to its coarse-grained representation then you see the buildup is much more similar. But it's important that you realize that some of the entropy has been replaced by enthalpy terms in this model which means that in principle the temperature dependency of the model is wrong and also in general the driving forces can be wrong. Recalibration of the parameters for specific temperatures is one way but it's not a very pragmatic way so typically one should interpret also the driving forces that you get out of martini with care. Another limitation is the screening of the implicit screening that we have in the martini model which means for instance that the coarse-grained model interactions are screened by a relative epsilon of 15 independent of the environment so whether charges are in oil or in water they are screened by an electric constant of 15 fortunately mostly they are in water so this is then realistic given the effective distance dependency of the screening in martini. But of course in reality as you capture in all atom simulations you have a difference in screening if you go from an aqueous medium to an apolar medium. There is a way to actually reproduce or to capture this also in martini and this is by using a polarizable model so we have a polarizable water model where you have two charge beats added to the central energy-owned beats they can rotate and they can kind of mimic the orientational polarizability of the four water molecules to which the coarse-grained beat is being mapped so using this polarizable water model and also together with PME you can increase the accuracy of your charged interactions. Okay then the last couple of minutes I would like to present some of the martini applications and I think these show you where the real strength of the martini model lies. So one is in brute force sampling so really you can attain very long time scales and you can observe processes happening spontaneously. So in this case we're looking at cyclodextrin which is an oligosaccharide oh sorry the presentation is interrupted so let's go back to where we were so cyclodextrin is an oligosaccharide that can actually extract cholesterol from a lipid membrane so this is shown here, the movie works again so here you can see that we initially have cyclodextrin surrounding a small liposome which in this case is face separated into this liquid ordered and liquid disordered domains there's cholesterol everywhere and the key question now is is this cyclodextrin being able to extract cholesterol and also does it do so from the disordered or the ordered domains and we see that in these simulations spontaneously the cyclodextrin binds to the surface and extracts cholesterol mainly from the disordered domains. Another example, a recent example of where brute force long time scales can be reached is in the photosystems tool complex where we looked at the exchange of electron carriers so here we looked at this beast of a complex, the photosystem tool that you find in the and the tylochoid membrane that is involved in the light harvesting process and the key player are these plastokino molecules that should be able to bind to the photosystem tool complex in their oxidized form and then if they reach their binding site and this you can see is happening in this simulation so it finds the entrance to the binding site the binding site is over here and there's actually or our simulations predict that there are three possible channels by which this electrode carrier can enter this photosystem tool complex so eventually it enters and finds its binding spot there it's stably bound in this oxidized state and then if you would do a few quantum calculation you can simulate actually the chemical process that is taking place but here we simply change the topologies to represent now the reduced version, the plastokino there we see that the plastokino as you expected to do immediately unbinds and leaves the complex to deliver the electrons to the next player in this electron chain so that was another example of brute force simulation then of course you can also use martini to go bigger and biggest and the current record is a paper that is recently published on the archive from the Hummer Lab by Vogler at Altres where they now have looked at membrane patches up to more than 100 million beats really connecting the molecular scale to continuum levels so they looked at the diffusion constant as a function of the finite size effects of the simulation box they eventually bridge really towards the continuum scale so more than 100 million beats is corresponding actually to 1 billion atoms so these are really giant type of simulations that are now within reach you can do this with martini another major application of martini is that you can increase the complexity of your systems quite easily because of this building block principle you can quite easily extend your library of molecules and in this case we looked at plasma membrane models that consist of more than 60 different lipid types differing in lipid head groups also different in lipid tails and setting it up in a completely asymmetric fashion the way that real plasma membranes are models or are existing and then you can do interesting analysis on for instance the cholesterol density fluctuations that you find in these complex mixtures this is actually ongoing research we're currently not sure yet what we're looking at here whether these are random fluctuations or critical fluctuations or maybe microemulsions and how this all connects to the idea that rafts are forming in real plasma membrane systems so increasing complexity the next step the current step we're doing is embedding also proteins into these complex membrane mixtures so to add another level of complexity you can look at protein lipid interactions you can see what kind of lipids are enriched around these proteins and again this is very fascinating ongoing research ultimately leading to really fully complex models of plasma membranes not only having all these lipid types but also hundreds of different proteins that are embedded in these mixtures so this is currently running on our computer clusters another thing you can easily do with martini is really high throughput studies so here's an example where high throughput martini simulations were actually used to predict possible peptides that can form nano fibers stable hydrogels so short peptides can form these kind of they can self aggregate but yeah there's of course already at the tripeptide level there's there's 8000 possible peptides that could be synthesized and if you go to tetrapeptides there's even many more so work by Fredericks actually a couple of years ago they simply generated martini models for all possible tripeptides and then predicted which ones would be potentially good hydrogel formers and then to do that they looked at the hydrophobicity on the one hand of these tripeptides and then doing short scale simulations looked at their great aggregation propensity and then they were identifying peptides that were aggregating quite strongly but still are on the more polar side because the more hydrophobic peptides of course all self assemble sorry here are the hydrophobic peptides they're all self assemble but they don't form these hydrogels but the more polar ones can potentially form these hydrogels and based on these simulations they actually predicted in this case a KYF peptide that indeed formed a very beautiful hydrogels and here in the simulation you can see also how these peptides self assemble into these kind of soluble fires okay another example where I think martini is strong is that you can combine different type of molecules together so here's an example where we combine polymers and lipid models together because these specific polymers are called steric, melaic acid co-polymers actually used to solubilize small membrane patches into nano discs and they can be used to extract membrane proteins with their surrounding lipids directly from real cell membranes but how this process actually happens at the molecular scale is not clear so we can simulate this at the martini level so here you see a movie showing how these peptides have a high tendency to absorb on the membrane and form these kind of small water defects while translocating and these water defects then grow actually into larger and larger pores and completely destabilize the membrane so at this point the periodic boundary conditions still keep the membrane together but you can also do a self assembly simulation and you can show that these peptides together with these polymers indeed form these kind of nano discs another example where we go away from biomolecules here is we actually modeling bulk heterojunction morphologies so these are mixtures of fullerines and polymers that are used in organic solar cells organic photovoltaics so we construct martini models of these polymers and acceptor molecules and what we do is actually mimicking the way experimentally these devices are made these compounds are dissolved in solvents like chloral benzene and then drop cast and spun so you get a nice film and then after drying the solvent out you eventually end up with this bulk heterojunction morphology so this process is mimicked here at the martini level so we started with a large box containing a lot of solvent and then step by step took away the solvent to end up with this final morphology and then you can actually compare this morphology to experimentally measured morphologies using actual microscopy images and you see we can capture these morphologies quite well of course now we have access to the details the molecular details of how the interfaces are formed using back mapping tools we can actually fully reconstitute the atomistic details and then use that for instance for the quantum chemical calculations of the actually exciton formation and splitting of the exciton into the charges okay the final applications that I want to show before the webinar will end is some new opportunities that arise from the martini 3 model so there's now the ability to look at a specific liquid-liquid coexistence of in this case ionic liquid and fish oil system where we see that oh this is the wrong movie that is showing okay this is a system showing a formation of co-asservates where actually you see that if you have mixtures of polylizons and polyglutamides or charged polymers in general under certain conditions they can segregate into two liquid phases one enriched in these polymers and one depleted in these polymers and this actually works now with the martini 3 model quite well and the other system that I wanted to show is this ionic liquid fish oil binary system so we have fish oil containing some of the polyunsaturated oil that is preferentially extracted being extracted by this particular ionic liquid so again martini 3 allows you to realistically simulate these kind of systems as well okay and then the final application also a martini 3 application where we're looking at a little benzene molecule that is able to find spontaneously its binding pockets in the T4 lysosine mutant where actually there is a pre-pocket as well as the binding pockets and eventually the benzene molecule that you still see happily sampling the aqueous phase so there's one benzene molecule in this simulation so it takes some time before it finally discovers that there's actually a nice place for it to hide away from the solvent into the protein and here it actually binds to the pre-pocket which is close to the final binding pocket but eventually it discovers this final binding pocket where it still has to compete with a tyrosine molecule that also likes to sit in this pocket from time to time and we'll see this tyrosine now trying to kick out this benzene again and taking the binding sites right here but then the benzene fights back and say look I'm the more apolar guy here so this is my place so this is an example that opens up I think also high throughput drug screening in protein ligand binding alright let me conclude I've shown you that the Martini force field uses a building block principle based on the 4 to 1 mapping scheme and parameterization uses both experimental top-down and all atom bottom-up reference data I've shown you a few examples of the Martini validation against a large variety of systems again either using experimental data or higher resolution models I've shown you that Martini is able to sample phase-page efficiently but of course at the cost of reduced accuracy I tend to call this semi-quantitative and of course there are certain limitations with respect to for instance protein folding and then at the end I show you that Martini can indeed be used for many types of applications involving both biomolecules and non-biological ones excelling in applications requiring either big systems or long time scales as well as high throughput systems and systems requiring a large complexity so with that I acknowledge people in the group so the current group members that will contribute to the ongoing development are listed here there's a group of key international collaborators Peter Tilleman, Ducamo Tecelli, Helgi Ingloff, Mano Melo and Xavier Perio that I'd like to acknowledge and also like to acknowledge the large international Martini community that contributes with their testing and coming up with new models and applications all the time that is very much appreciated so with that I think I give the screen back or no I keep the screen I think but I think this session is now open for questions and answers Yes, thank you Silvan for the great presentation and everyone if you have any questions please use the Questions panel Could you please keep the slide? Yes, that's good You can use the Questions panel to type in your questions and I'll give you the microphone to ask it or I will ask it on your behalf I have one question So Martini has also been tried in multi-scale hybrid mode, right? How well is it working? Is it possible to mix with atomistic maybe on different scales? Yeah, that's a good question We do have some multi-scale models where you can directly combine Martini and atomistic force fields in one simulation the so-called hybrid simulation this can either be done in a static way where you have a certain predefined area in full atomistic detail surrounded by coarse grain beads much like in the QMM approach or this can be done in an adaptive resolution way where coarse grain molecules if they enter a central zone also change their resolution So this works for... so we've demonstrated this in a number of papers that it works for certain key systems but the problem a little bit is that in the current implementations the speedup that you get is still very limited so at the end you could as well run the system fully atomistically so this still requires some optimization of the way that these algorithms are being implemented in the major simulation codes but in principle there's ways of combining atomistic and coarse grain models directly Thank you So we have a few questions There is one which just says polarizable Martini, a question mark So I guess the question is do you plan to develop a polarizable model? Yes, in a certain sense we already have some polarizable version of Martini the Martini 2.2P model features some polarized polarizable side chains that in combination with the polarizable water model gives you a much more realistic description of the electrostatic interactions There's also ongoing efforts to introduce polarizability along the protein backbone that eventually would also allow us maybe to simulate protein folding again in a more realistic way so there's definitely ongoing efforts to introduce polarizability into the model but Martini 3 is standard Martini 3 will be non-polarizable but also there we will have flavors coming up that have polarizability included Thank you Next question is from Zeynep who is asking, I'd like to know when simulating proteins is it important to use only the enforce field or can we get as good results using Martini 2.2? The Elnidin force field for most proteins we use the Elnidin force field smaller peptides like single transmembrane helices you can also stabilize with the appropriate dihedral angles but for the larger proteins typically we do use an elastic network or we're also now exploring more and more this Go Martini approach actually you can also see partial unfolding of the protein or peptide involved what you also can do so you don't have to constrain especially for larger protein complexes you can apply the elastic network only to certain domains so you can have domains that you want to keep stable you stabilize with the elastic network but the motion between domains you can cut so then you can have full motion between the domains the example I gave on the channel gate opening of this mechanosensitive channel is an example where the individual domains could move with respect to each other and give rise to the opening of the channel so this gives multiple options of using these elastic network models Thanks Another question is by Ash Tosh Tripathi who is asking whether it's possible to look at conformational changes in the protein where there is a change in the secondary structure for example partial unwinding or coiling our helix Okay again this comes back to the same point so with the standard elastic networks you cannot do this because then the secondary structure is fixed but again you can either tweak or cut bones in the elastic network that allows you more flexibility and then if you can compare this to atomistic simulation you can try to capture the same type of fluctuations as you see in the reference atomistic simulation and this goal martini approach might be a way where you can also destabilize the folded conformation and go from unfolded to a certain folded conformation with the martini model Alright There's another question by Hülso who is wondering why the POPC POPC for suit has changed between 2.1 and 2.2 and there is one extra, one less atom actually in the new version The reason is that we found that the original 5-bit model for the POPC produced bilayers that were way too thick so the POPC is a level of if you count the number of carbon atoms it could be presented either with a 4 or a 5 bit is kind of one of these in between type of lipids so more extensive comparison to atomistic simulations and experimental data decided to go to the yeah this so much shorter 4-bit version that an R view performs better than a 5-bit model Here's another question what happens to cost drain models developed with martini 2.2 whether they will reproduce the secret behavior if you start using 3.0 with them I guess this is tested already Yeah, so we try to test the 3-bit model we of course recalibrated the whole interaction matrix so in principle the behavior of all systems will be affected most of the topology so the B types that are used for all the lipids for instance and also the protein side chains are not changing but their interactions will have changed so we're doing a lot of tests to make sure that the system that behave well still behave well at the martini 3.0 level but of course we cannot run every single system that has ever been simulated with martini and for sure whenever you improve on a force field there will be hopefully in general an improvement but there will always be systems or areas where you actually will see that the system becomes will not improve or even will become worse this is unavoidable again this will have to we have to see if this is the case also with martini 3.0 but we hope that for most systems that behave already well remain behaving well Another question from Matthias Marcado is how sensitive is martini to electric fields I think this was partially covered in the talk Yes, so electric fields there is charges in martini so everything that bears a charge ions or head groups of lipids or charged amino side chains will respond to an electric field of course their response will be much more realistic if you combine this with a polarizable water model and so there we have for instance shown that with a polarizable water model you can for instance study electro operations of membrane electrostatic field across a membrane it is known that this creates pores in membranes and with the martini polarizable version you can actually also see that happening in the simulation so there in principle you can play with electrostatic fields and you should see at least qualitatively a correct behavior appearing There is a question by Joshua if martini 3.0 can combine less than MB models in a single simulation could you use a smaller bit at the end of your MB to distinguish between example 12 and 14 C lipid tails Yes, that's a good point so with martini 3.0 now we have also fully recalibrated and integrated these tiny bits into the model so that gives you also more ability to fine tune and indeed distinguish between 14 and 16 carbon tails so indeed this opens up opportunities in that direction Thanks There is another question whether you would recommend certain software for doing Monte Carlo simulations with martini Well there I don't think I can recommend anything from any Monte Carlo so I'm not very familiar with Monte Carlo software I'm also not sure I think people have tried to do Monte Carlo on martini I mean I see no reason why martini should not be able to be sampled using a Monte Carlo algorithm but yeah so I wouldn't be able to recommend specific software for that Yeah in principle it should No There is a question which I don't understand completely but could you give insight in regard to the usage of martini to study coaservate Coaservate Yeah so coaservate this is one of the last examples I showed so with martini 3.0 we have now been looking at coaservate systems very many different conditions and it seems that we get quite realistic behavior so indeed we can see this phase separation of liquid phase separation where one of the liquids is enriched in these polymers that give rise to these coaservate systems and a region which is depleted so this is something that we're definitely trying to embark on more and more in the future because there's many interesting questions to be answered because nobody really knows what are the driving forces for these coaservate coaservate formation and what is the what partitions into which of the two phases so there I think martini can play a nice role Thanks and we have a number of questions but one last question because we are already quite over time are there any improvements that are being made in the representation of ions and their iterations specifically for calcium ions Yes that's also a good point so standard martini works quite nicely for standard ions sodium chloride performing an overall screening that these ions do also in reality but yeah so far the ion models have been quite crude a good model for calcium is indeed not available yet but with martini 3 also there we have much better interactions for ions including the divalent ions so there also we found that using the small or the tiny particles sometimes can also improve the description of these ionic solutions so yes you can expect also better ionic parameters with a new platform martini 3 Thanks and there is still a few questions but since we are quite ahead of time I would ask our audience to continue maybe using the support forms of martini that you showed us earlier in the talk and so could you yes so in the next few weeks our webinar series will continue with presentations by two companies one is from Bikki Technologies which will tell us about finding a trade-off between speed and accuracy during the simulations of protein ligand binding models and we also have a tent of May presentation by Novartis on a NMR guided locking of protein ligand complexes so with this we will finish today's presentation I want to thank again Sivartian for the great talk and see you at the next event Thank you for seeing everybody Thank you