 Okay, here we go. So we have VS Gapsis visiting from Bertha Groups Group in Bifinion and this is a little bit of an experiment. It's the first time at least here recently running a seminar where we're doing a live stream slash virtual seminar at the same time as doing one in person. So please shout at us or drop something in the chat window if you run if you have any problems as we're going and we'll do our best to resolve them. We have about 10 or 12 people here in the room and it looks like about 13 people online. Anyway, VS has done some really nice work that I got to hear about earlier this year in Germany and on applying, finding free energy calculations of various types on a fairly large scale. So I'm excited to hear about that again today and bring that to all of you. So please take it thanks for the introduction and for the delay, let me start immediately with this topic. So today I chose to tell about a particular application of our approaches that we developed in Bertha Broth's lab and it's true that we usually use non-equilibrium free energy calculations for various systems. So we have applications for amino acid mutations for DNA nucleotide mutations but today I will concentrate mainly on these approaches for the ligand modifications in the realm of relative free energy calculations. But later we can also have some discussions on those other types but the core of the presentation will be about one of these topics. Yeah, that's exactly what they just said but it just illustrates that bit more in depth. So what we're working on in Bertha's lab is one of the areas is development of the methods that we term BMX. So this is the software that we're developing and it is capable to prepare the input for the simulations for the chemical free energy calculations that will allow you to do something like mutating amino acids and in proteins to assess protein thermostabilities or mutate nucleotides and DNA and the modifications which is the topic for today. So yeah let's dive into the ligand alchemy how we do this. So we just a few words about them more the technical side of this. What we have is a framework which relies on basically on three different modules. So let's start with the first one. Firstly we'll need to identify first what we will want to do so to calculate relative free energy between given two ligands. We would like to more make a more between the two of them for that we'll firstly identify atoms to be more and we will build a hybrid structure anthropology for the two of them and we have also the third module there in the upright which actually will help us to navigate in a given chemical library to suggest the ligand pairs to be more. So firstly how do we do this atom identification for mores? We have two ways to go either alignment algorithm could be used or a maximum common subscription. They both have their pros and cons and this is yeah it's nothing new we're just using algorithms that are available. You can ignore that somebody's not needed and I'll try to fix it. All right and yeah what we have is a quite now yeah an involved scheme that we do not need to go in detail but basically there are those two parts either we take two ligands if they match quite well geometrically right you can simply super impose them and say hey we will now make a scaffold or maybe and the rest will be done. This is one way and another way is for example maximum common substructure matching which is a topological matching right and there might be many caveats along the way that's why these parts are so involved right there one needs to have many checks along the way and for example what could happen if you just use a maximum common substructure matching and you have two very different molecules in in their very different states maximum common substructure will say hey they're completely identical there is a very similar scaffold and it will suggest you a mapping that is yeah just as an example here those atoms to be morphed into one another while alignment would say only a fraction of these atoms matches and depending on what you actually have are there chiral centers that are involved or are there some polar groups that you maybe don't want to morph we have encoded a number of empirical rules so it's very similar to what it has been done in Davis lab and previously well in their own implementation their own set of rules so I think it's a very much a complementary way to look at that and afterwards we so we have now identified which core which core atoms should become one we can build the hybrid structures topology so this is that's just a technical part which I don't have any illustration for something happens under the hood and the third part so this is also a very similar approach to the building to the low map approach to build graphs we just have it in in our own software in our own formulation so how we formulate this building procedure is basically given a small chemical library of compounds that are depicted here we can now construct some similarity measure based on how many atoms are being morphed and how many of them are to be become dummies we build a similarity or distance matrix and based on that we can start constructing some graphs of interest either minimum minimum spanning trees or or in introduction of redundancies can be done so this is something quite established in the field and we yeah we just do it in PMX as well all right so these these are all the technicalities that we need to know how how we then validate our approach so we look we looked at the following thermodynamic cycle so the question was is our implementation robust and do we do everything properly so to do that we looked at the cycle that's here we calculated for for given two limits we can calculate absolute solvation tree energies for ligand a and for ligand b and on the other hand so these two paths these two vertical paths would not require any hybrid structures and topologies there's simply the standard code in simulation package we were using grammar by the way I didn't mention but this is not not so important at this point and we can also build that hybrid structures and topologies with our tools and if the everything is done correctly right these two the difference between the vertical branches and the horizontal branches should come out at the moment and of course then we need to test this on on a large set of ligand morse so we use the there is a library the free solve database with all the parameters there and it's it since it has more than a 600 set of 600 ligands we could construct a huge map with very different very different perturbations really pushing us to the limits so because it requires all of them to be morphed into very different chemistries and then yeah we we get a very close agreement it's not perfect so here I'm just depicting on the left there is our relative versus the absolute two branches versus two branches the differences and then most of them fall within the uncertainties of one kilo calorie or so and but this also serves us as a very good set to improve on those rules that we kept having sometimes we would see an outline like I'm here intentionally showing that we did have some outliers like one percent of those 600 edges and they were they always revealed us something unexpected like I mentioned some chiralities yeah which should not match those atoms to be morphed they are chiral centers atoms around chiral centers and yes things like that and then we can prove and further our our sets of improvements all right so this was our validation that's it's done with this part which is maybe a little bit boring or so it's nothing applied so far but then we took our tools to to a real real world cases we looked at them almost 500 ligand modifications and I think 13 or so protein ligand systems and we for all of them we calculated the following cycle where we wanted to know what is the difference in the energy so a double difference in the binding in the binding affinity upon a ligand modification here yeah before before going into the results I would like to know that we do not use the classical approach well state of the art approach in the field of chemistry which would be free energy perturbation all the results we will compare them to this FEP protocol and FEP just to get everyone on the same page is the protocol where we would along the chemical coordinates we would discreetly distribute lambda points and simulate in in equilibrium conditions at all of these points and get free energy difference between the neighboring points by means of most usually benefit acceptance ratio estimator but we usually do not rely on this and in this application we also use the non equilibrium approach which works as follows so we are simulating we're performing two long simulations in states a and in state b where one state is different a another is modified and gather snapshots from those two simulations and start very fast transitions from a to b alchemical these are transitions are of the order of 100 picoseconds or so each of them and we have many of them each of them is out of equilibrium of course so none of these work values that you gather would be can be considered as free energy difference but we can rely on a consultation term and get the free energy difference back out of the work that we distribute all right so we have that and yeah as i mentioned now using all these tools we looked into a number of diverse systems so some of these systems are in the quite well known data set already explored and yeah gathered by Schrodinger corporation and it was published in i think 2015 and jacks which so it's known Schrodinger's jacks data set but it's not only that so those would be i think somewhere starting from here or so but there are other five other sets that we added in addition so we have now quite a large data set and we can immediately let's immediately go to the results how do we perform so now on the left here we will look at the FEP plus so this is Schrodinger's data of the art FEP based approach using their own proprietary force field and on the right there will be different incarnations of our calculations those will be either using that or using charm's charm and with the CGNF that or we will be using consensus in our in our work with with other biomolecules so amino acid mutations and nucleotide mutations we have previously identified this peculiar feature that the force fields tends to certain force fields tend to point in opposite directions from the experimental value this is just an observation there's i am not claiming anything but the rigor is there simply you would make an error in one direction in gap but in the other opposite error in CGNF and just averaging the results in terms of free energies and they gave more accurate the description of that of the more accurately matching experiment values all right and how we're looking at this on the top we're looking at the average unsigned error so the lower number the better and the lowest Pearson correlation coefficient is the higher number the better and each number each symbol here is an individual repetition of the simulation in each circle we use 60 nanoseconds per delta g right and we have like 500 of them of these delta g that's been average over and right so in terms of average unsigned error our approach actually performs equally or yeah gap is quite close charm is quite far from the from the Schrodinger's calculations and we get back to identical numbers as Schrodinger has and in terms of correlation it's very similar so gap is close charm is a bit further and in the consensus we're closer but yeah it's it's a yeah something within the error bar slightly lower correlation right so we can go a little bit deeper into this as I mentioned there are actually three datasets so the first or two merge datasets one is from Schrodinger's and Jack's dataset another is our additional dataset this is the same picture that they showed for now if we look at what what can we see here is that actually if we were to look at the original Schrodinger's dataset we would be doing better than them in terms of average unsigned error but in the correlations slightly lower and there's one more point here which is now a bit transparent here I'm trying to show you here the mouse it is the value that was reported in the original in the original publication this Jack's paper and why is it a different number well it is actually using an older version of OPLS so we can also see how this OPLS improved because for our calculations we used the latest one so did it improve well not in terms of the absolute agreement of absolute unsigned error but yes it did improve in terms of the capturing the chance and we'll nail down but which systems actually showed us gave us better agreement and our datasets in terms of our dataset well we didn't do ourselves a favor here adding those additional datasets we actually do a bit worse than than FEP plus in correlation where there's everything is equivalent but in absolute terms a little bit worse but yeah let's let's see where these differences come from so we can break down everything by the by the system right now we had 30 systems and this is our consensus result so we have lots of scatter plots and we can analyze all of them for a long time but yeah let me maybe just summarize a few points that you can see so there are systems that behave very well like this one all the points are blue so we are blue is mainly where we are we didn't want to calculate anything or so from our from the experiment but there are systems that really behave yeah they're all over the place like this on sale one it's a it has outliers both in the small value range and the large value range there are some interesting systems where we have all the points are blue but we have zero correlation or so yeah well and this is all because of the very low dynamic range I think I'll go to this in the next slide or so but mainly where where the errors come most of the time and this will be summarized further but the basically they are in the lot for us it's in the in the areas where we have large free energy differences so we have an edge where an error where a free energy difference would be large we do not capture it very accurately or less accurate than the smaller change and I think this will be now very should be summarized here yeah exactly I looked into now I broke all of these I broke now all of these data points that they had into into the following ranges for the free engine difference between the two ligands that we are looking at is within oh I just lost your slide yes that's it okay so now we'll not touch anything all right here we go again okay so what I was saying this is exactly illustrating simply what I was trying to tell so many words but the basically the small difference in the small free engine for the small free energies we captured quite accurately so we have a better agreement with experiment or smaller error in terms of yeah then then the FEP plus approach but when we go to for one to two kilo calories differences we still capture them very accurately it's all comparable but the differences start appearing when we have large differences so two to three calories or even more of course the statistics gets lower we don't have so many edges there but yeah it's really that's where we are doing slightly worse but this is also in some sense wouldn't use because because we can frequently this probably is happening because of the sampling issues we're not using any enhanced sampling for this study we in comparison to FEP plus which does and maybe there are also some dependencies between the methods we could discuss this maybe in the later in the discussion but yeah so it's we understand at least this effect just understand now if we this is a very busy slide but just to briefly go over it it is the again average unsigned error and correlations for broken by the system for all of these methods but only the only message that I would like to bring here is that it really is diverse so it really is it really strongly depends for whatever method you look at either it is driving a safety plus or for our approach it will depend on the system very strongly so correlation can range for us from zero to one right but the same goes for any of these points either red or blue or so in any general gap performs better than for every system for most of the systems better than charm and the CGNF and the consensus is outperforming those both of those and yeah what happens to sometimes that is peculiar is again this issue of having very for thrombin inhibitors for example very low error here so we have almost the low one of the lowest errors but we have no correlations so and why is it well it is because of the very small dynamic range of the delta delta g values that's what I was trying to show here so simply showing all the here it's again all of these systems we can just concentrate in this case of thrombin inhibitors we have experimental value range plotted as a histogram so experimental values differ between one and between minus and plus one kilo calories per mole our approach recovers that range very well so this gray line I don't know if you see but it's really overlapping with experiment and the Schrodinger's approach likes to exaggerate those differences so they do have larger absolute errors but this actually plays and they have been capturing them the trend so yeah that's well this just comes out as an effect of this of the slightly less accurate absolute correlation and we also so we are talking a lot about accuracy and precision and let's have a little bit of those are looking to that now if we look at accuracy only so I flooded here now unsigned errors for all of these 500 ligands for all of these approaches either with tubs, genotape, concerns and master and we already yeah I mentioned several times that the accuracy on average is similar and the distributions also look very similar so yeah there are many curves but they all overlap more or less well so all of these approaches they approach the same agreement with experiment but if we look at the precision so the uncertainty now the distributions become different the distributions are yeah those curves and the bars are the mean values so if we can just during our eye we just let's look at the bars and my strategy B plus is very certain in its answer so if we repeat it three times I didn't mention all of these all of these simulations or every delta g value was obtained by repeating the simulations three times at least so that all the uncertainties are actually not just uncertainties of an estimator but actually uncertainties of the sampling of the configurational aspects so if we repeat three times the FPP plus calculation is still give almost identically the same answer within one kilojoule per mole or so if we repeat now with any of these our approaches we get a larger uncertainty between well around two kilojoules per mole or so on average so this is really substantial and we are still looking into this but so far it's just an observation but the well in defense of these approaches we just looked at the accuracy so we know that accuracy is identical just one approach things that is very certain that this is correct answer another approach is not so certain but I'm not claiming that this low low precision is a valid and all right so now for the for the next part of the talk let's have a look at several systems to see whether we can understand what what is driving these differences in accuracy like we saw so for the first one I looked into ETP1P this is a quite an interesting system it's a phosphatase so it takes a piece phosphate away from tyrosine and it participates I think in one of the signaling pathways then but for us it's what is interesting in this case it is one of the members of the Schrodinger's Jagdstatus and it is it has a catalytic system in which it is known to participate in this cleavage mechanism and it's known in that in apostate this system is deprotonated so it's oxidized state but it's not just in such a form like this in apostate it actually is shown from a crystallographic structure that should make even a covalent bond with a with another neighboring serine residue so something strange there but all right we know that it should be deprotonated in the apostate although the further phase is not known but now what happens for the test case of the ligands that we're looking at well these ligands make this is a crystal crystallographic position both of the ligand and we know that the distance is quite small so it's on the edge of making a weak hydrogen bond potential and if we just run some predictors empirical predictors this was of course carboxy group would be predicted to be deprotonated and cysteine would be predicted to be protonated so all right so we have these predictions however in the original Jagdstatus set what the Schrodinger people modeled was the cysteine deprotonated cysteine so they had have no proton there and the result so now it becomes a bit in quite entangled but let's have a look chronologically at all the results so this is the first on the left is the original Schrodinger's publication in with the deprotonated cysteine so cysteine the charge minus one and the values that they got right so the standard average cysteine error and correlation they're not particularly good but well in decent regime and there are some outliers there right now in the new version new release of their force field they removed completely all of these outliers and but a perfect agreement almost almost perfect agreement with experiment now if we were to follow this logic that we kind of new kind of convinced ourselves that there might be a chance for my system to be protonated we can protonate it and and calculate I'll come to John today on the screen it looks like he's still okay but why don't you okay at least a moment we have your slide up here still all right okay so what we have then is the is when we protonate the cysteine we have of course loss of accuracy so the force field only tolerates the deprotonated cysteine so it was really almost seems that it was made to be exactly like that now for our approaches we first start with the deprotonated cysteine and we do have now comparable results to the early Schradinger's approaches then if we make an artificial construct of deprot of just removing the charge no addition of a proton but just neutralizing cysteine just something intermediate we get a slightly better result and then maybe even better visible on the correlations and if we protonate cysteine we get even better accuracy so really for us the best accuracy agreement with experiment would be if the cysteine was in its protonated state which we have predicted but so this simply shows how one could modulate well fine tune the force field to reproduce well the values that are in your data set without yeah well by driving it into some certain direction which is not necessarily a carbon stone that this has to be steel group has to be protonated and the yeah it's I'm not claiming that it's really like that that it's up to that sort of use this thing that should be there but I'm just saying that one oh no we have another maybe we should try to join one more time because it looks like if I do it for my laptop we'll get a mirror image of the video so people can't read text I'm guessing that's internet connectivity problems if I give you a LAN cable like an ethernet cable plug for it I have an adapter if you take usb-c yeah okay so why don't you work on reconnecting if even if you drop out really quick and I will grab an adapter I'm gonna pause the recording it'll be right back right back online while I grab an adapter from the store okay I hope just hopefully we're coming back online now I hope it will be better now I can switch Wi-Fi networks okay so you just finished this slide right we've finished here and I wanted to go over just several more systems because they show some interesting behavior as well so this is another case that we looked at and found something interesting so here we used GA in this example to calculate some free energy differences for galactic and in one case for example here let's have a look we have quite an inaccurate result this is already the difference this is not the change difference but this is the error that we make from the experiment so this is quite inaccurate and what is this so we could immediately make a conclusion that yeah all right so one of these so these are all relative changes right so one of these chemical groups is probably parametrized wrong right is it this methylamine or or or or yeah another one well we look at the or methoxy group then we look at the let's say we probe this this methylamine with the replacing it with the methylamine and we see a small error so all right then this must have been a toxic group that was wrong but then we got probed with the by replacing it with the hydroxyl group and no it's also no error right so those so what it seems that the force field was able to capture correctly the populations of the intra chemical groups but not in right so there might be some very yeah and this is difficult I think to both well now we captured it but it's also difficult to know in advance how to parametrize when parametrizing the force with one doesn't really care about these well it has to be an absolute absolute pre-energy that is correct but it seems that there might be someone expected the caveat so on the way and what I meant for the last system I picked one to look at where we actually what I mentioned to illustrate the effect of force fields pointing in the opposite direction from the experimental value and the cement system had a more than half of all the calculated preanges were going in different directions from the experiment it's a business slide but what they would like simply to show here is that yeah we're looking at the at the ligand which is here and we're only replacing the substituent with many of them and some of these substituents really let's say let's look at one of these codes for 2015 so it's this one for whatever reason none of these others but for only this one in four different cases it decided to go to be overestimated in triange and gap but not but underestimated in CGNF of course this doesn't directly give me a clue which exactly parameters so are those the charges or are those the the idrills and what to fix but it simply tells us that there are two sets of parameters gap and CGNF that needs to be yeah they they need to be I'm not saying that average but taken into into account maybe also maybe this brings us to the limitations of them of the classical single point charge models of charges and force fields right maybe one needs indeed needs two sets of parameters to represent certain chemistries and also it's quite peculiar because these substituents they just looking by eye they look quite similar and there is no at least for me an immediate reason why one of them should always be overestimated and another another others perform just fine and other comparisons right but this is just these are just the observations from all the scan and yeah here it's already the summary I already spent more than half an hour on this so what we talked about was relative free energies and the application of those but if we have time and if anyone is interested in I also have some backup slides on the things that we are currently working on on the absolute changes and also frequently I get the question so and we also looked into closer into the comparison of the equilibrium free energy perturbation approach and non-equilibrium to see how well they perform but yeah maybe we can start some discussions questions in discussion so far this is uh Mike Wilson can you hear me yeah oh okay yeah really nice work thank you I've um question about the how you did the non-equilibrium stuff so was it essentially I mean you make windows and then partially equilibrate each window in effect or is that an alternative to bar and bar or is there is is it almost like a sort of a pulling as it were through parameter space I would say the latter we did not do any windowing or so those were really two equilibrium simulations so without just standard mb simulations at the end stage and from them I would say it's an analogy to pulling so taking one state and pulling but along the alchemical coordinates not the physical coordinates and really pulling very fast to reach the other state but not to reach the equilibrium of the other state but simply to reach the well Hamiltonian of state b completely physical Hamiltonian so and then so you do some in one direction and then some in the other direction basically and collect statistics yes exactly we do a bi-directional pulling and collect the well distance in terms of work values those work values are three energies but with some dissipated work components so due to friction or simply due to the fact that we don't end up also in the equilibrium state where so how many how many do you do in each direction all of these were done on the order of hundreds so let's say a hundred in one direction and a hundred in another okay and then you said some at 60 nanoseconds so was that spread over the hundred or is that this is the accumulated time in total it is 60 nanoseconds of time spent per delta g so i think for the equilibrium run so all of these it's important to start those of course one equilibrium pullings from the equilibrated ensemble for the equilibration of the ensemble only used six nanoseconds per two states of course so all the rest was then done for pulling okay thank you hi this is Dave I have a question following on to what Mike said which is so how do you calculate the uncertainty in the work done during this non-equilibrium process or what is the typical value of the uncertainty and then since you end up doing it multiple times in both directions do you come up with some summary statistic that gives you the uncertainty in each transformation yeah that's a good point so we calculate uncertainties for both in two ways and combine them into one estimate of uncertainty so first uncertainty comes from the from the estimator simply we rely on the maximum likelihood estimator for the three energies so once we collect all of these work values we then can use something that is essentially benefit acceptance ratio for the non-equilibrium three energies that's a it has equivalent form of that and then of course we can bootstrap uncertainty by simply bootstrapping the work values to recalculate this estimator of course this is just the uncertainty that is yeah i call it the uncertainty of the estimator of the estimation but we repeat everything as i mentioned multiple times all of the starting from scratch all of the simulation both equilibrium and the non-equilibrium part several times and this gives us another uncertainty set up uncertainty and uncertainties which is also covering well some i would call it sampling uncertainty so if we were to be driven to a different phase space region by just by the stochastic nature of md we would actually also capture a different three energy difference right so we all of these uncertainties that i was reporting combine both of these factors yeah that makes sense thanks i'd love to hear about the absolute in the equilibrium versus non-equilibrium briefly if you can manage it yeah yeah sure those are only a few slides and oh i'm sorry i had to acknowledge people that work on this and the findings and everything so here's a very quick acknowledgement thank you now yeah let's have a look into the absolute because this is a bit of a different beast right we then we need to really think of the problem in a slightly different way we need to make a large perturbation it's always most frequently will be large it's larger than in the relative case convergence will be slower slower will require restraints but yeah it is possible as we'll see to do it also in the non-equilibrium way so we started with and looking into this from starting from the work is done by a material material day who looked into the absolute changes a few years ago based on the equilibrium approaches and yeah this is just a standard cycle which also looks slightly more complex than for the relative changes but the yeah we don't need to go into the detail of the cycle this is i think there is standard way simply there is an additional component of adding restraints to the decoupled state of the ligand right and what Matteo looked at some a few years ago was promos binding to different promo domains so he took 22 different proteins and one single ligand and docked the ligand and later calculated the free energies to those different proteins so basically probing the selectivity of this ligand to those different promo domains and he got very good agreement with the experiment with the equilibrium FPP approach but yeah it required quite a bit more sampling to reach the convergence it took a 600 nanoseconds in comparison to 16 nanoseconds per single delta g value and these are his results I just recovered them from his publication and we did exactly the same now this is a bit more data on the slide we did exactly the same with a non-equilibrium approach well first of all on the left it's exactly the same what I just showed from the S publication the same plot and then we recalculated everything just to have the same data in in our setup with FPP equilibrium approach we got slightly worse agreement there are some statistics given so every some some some estimations of uncertain of accuracy is given so every sunset error let's say is one and a half kilo calories per mole which is slightly lower than usually we would expect from the relative changes what we got is above two kilo calories per mole and it's a bit difficult to say for sure but the one difference the main difference is actually that Mateo was using replica exchange so enhancing the sampling between between the discrete states by making the transition between the lambda states we did not do this in this case and yeah but also we can see that in other metrics we this approach did not do worse actually did even better we get an even better correlation in the within this data set and this is the non-equilibrium TI this approach of non-equilibrium and we get very close within the uncertainty of the average sunset error and we see within the range of uncertainty in comparison to the previous approach and again slightly better in this case correlations whichever correlation was in the path so yeah this is our investigation into the absolute changes and yeah just one more additional slide we can also probe probe the convergence how quickly it happens how quickly the results convert here I'm looking at the RMSE and Pearson correlations for different approaches so let's concentrate maybe on one of these RMSE so this is our reference what Mateo got from his study and actually it's quite quick that we could reach this value we would get already the same accuracy just by having this half of this simulation time and it seems that the FEP takes a bit longer to converge of course this can be facilitated by some replica exchange or so but without any enhancement it seems that non-equilibrium converges a little bit faster but of course there are other lines that you can see they are just more for the amusement those are the single one one directional pulling so based on Jerezynski's equality in that in these cases and those could be either directly using Jerezynski so which would be then biased estimation of the given the final sampling of the pre-NGs I think these are the green ones and the the or or the darker ones and the other two are based on the Gaussian estimation cumulant cumulant expansion of the Jerezynski's equality those should be unbiased however they suffer from severely from other artifacts so yeah I would not recommend going with that one directional pulling experiment right and this so this is on the that's what they had on the absolute green energies and another few minutes I could spend on that a little bit on the preliminary results what we have for the comparison of the efficiencies or we can just yeah I think there are a few minutes okay so really a few minutes so there have been a few studies now recently that aimed to do that and we just wanted to have our own approach because we wanted to address three main main issues that usually arise in these comparisons so first we wanted to have a large number of realistic perturbations those that would be we would frequently encounter in our realistic applications for mean ice imitation for example or ligand modifications not not just taking one particular example but many of them then we wanted to set equivalent conditions and sampling time so that's really not that we would be running for 10 nanoseconds one approach another for 100 nanoseconds and then somehow try to meet them by extrapolating the result by assuming that samples are uncorrelated and then uncertainty that also by a square root of n or so but no really having equivalent conditions and the third part was a known target value so if one approach converges to one value another to another but we don't do not know what the true answer is it's a problem and we started constructing these closed cycles where we perform in the single simulation box for two systems that are restrained in far apart from one another far beyond the cutoff the real space cutoff value we start perturbing one system into another but we have we also make exactly the same perturbation in the opposite direction and this way construction needs to give us a zero value so we have a target value and immediately let's have a look at we we probe many of them first we started by probing all the amino acid perturbations in the relative pre-energy calculation scheme and those are yeah this each of the curves contains some maybe I think about a six microseconds of sampling time and each of them on the x-axis is a sampling percentage so how much time we spend to sample to get this number and the delta delta g which we just agreed that it has to be zero is given here on the y-axis so if it is let's say here going to zero then this protocol is giving us very very good convergence these are simply different protocols that we could probe for then an equilibrium guy we could spend more time in the equilibrium state we could spend more time in the transition state we can run then more transitions but the but the one requirement is that each of these curves then will spend I think 60 nanoseconds in total per run and we run it multiple times I think 10 or so times and we can compare them equivalently to the equilibrium FEP and we see immediately that there are some some problems these are again different protocols give some some of different values you can use more in this case we can use either more windows or fewer but longer windows or we can use a separate splitting of pulamic and linear zones interactions so I'm just showing a selection of those we have I think 10 protocols for one and 10 protocols for another these are the best performing ones so we do see quite severe problems when it comes to charge changing well the the box conserves charge but the mutation is charge changing so yeah this is one example then we looked at other set of in this case ligands so these where I mean I says now we are looking at ligands those are really the salvation pre-energist salvation in one direction salvation you know or decoupling in one direction coupling in another and it's a very similar picture that the non-equilibrium converges faster than equilibrium which sometimes is really giving some of results but again I have to note here that all of these equilibrium are without any enhancement of something so we just have had equivalent conditions for both so yeah these these are quite preliminary results in terms of right and I think yeah I don't have any further on this thank you okay cool this is this is really good I'm looking forward to talking with you more offline so thanks everybody for joining and thank you for such a great seminar we really appreciate it