 Hello everyone and welcome back for some of you to our virtual workshop on best practices in QMM simulation of biomolecular systems. Now, just to reiterate and remind why we're doing this workshop, we know that there's a significant interest in using QMM simulation of biomolecular systems. However, we also know that some of the key challenges in doing this are in choosing a suitable QM treatment, for example, the choice of DFT functional, the choice of basis set to choosing and tuning the QM region size, and also choosing a suitable reaction coordinate. And we know from surveying people are interested in QMM that about half of them were significantly hindered in using QMM simulation because they were not certain how to choose suitable QM treatment. And we know that we need to be careful when simply using QMM because it's dangerous to use particular choices and functionality implemented in software as a black box, i.e. without knowing the suitable applicability and the regime in which these different approaches can be applied justifiably based on the underlying quantum chemistry. So this can lead to bad science or even worse, rejected publications. So, why we're having this workshop is to enable people who have significantly experienced in using QMM simulation for biomolecular systems to highlight various aspects of best or good practice that includes the potential power as well as the limitations of some underlying approaches. To highlight tricky but nonetheless crucial aspects that are involved in actually practically doing this kind of simulation to warn of dangers and pitfalls that can come up and to illustrate these in the context of concrete examples based on their own research experience. And we expect that this will benefit people who are less experienced by providing insights into how to go about determining a suitable modeling approach in particular choosing rising to these key challenges of choosing a suitable QM treatment and reaction coordinate. Thereby developing also robust simulation protocols that will lead to well justified and meaningful results. And it was also worth saying is that a lot of these questions are actually preconditioned to choosing the software package one might want to use depending on what functionality is implemented. So to introduce today's speaker Maria Konova is based at Lomonosov Moscow State University and also heads the molecular modeling group at the Version Academy of Sciences Research Center for Fundamentals of Biotechnology. She's interested in molecular modeling of enzymatic reaction mechanisms as well as photochemical processes and proteins. Maria's talk today is going to be on the validation of DFT functionals in QMM simulation. And this clearly given what I've outlined as the workshop goals. This is clearly a very, very relevant topic. So I'm expecting this will be a very interesting talk. So with that, I will hand over to Maria. Good afternoon everybody. First of all, I want to thank the organizers for inviting me. And it's really nice. I mean, nice set of webinars that deals with the very specific field of QMM simulations of biological systems. And it's a great pleasure for me to contribute these series of seminars. So I'm going to talk about the very particular field of enzymatic reactions that I initiated by the nucleophilic attack. And yeah, actually, nucleophilic attack initiates wide range of enzymatic reactions. Those are mainly happening in hydrolysis. It's class three, according to the enzyme classification. And these hydrolysis act on Easter bonds, peptide bonds, on CN bonds, on the other, well, of the other names. They act on nature than peptide bonds and also acting on the CC bonds. Importantly, these hydrolysis are able to cleave covalent bonds that are inert in aqueous solutions. So these compounds are do not degrade in aqueous solutions, but after being into the active side of protein of the certain protein, these compounds readily degrade. Well, in the lower part of the slide, you can see the scheme of these nucleophilic attack. So there is a covalent bond between carbon atom and X atom, X is nitrogen or carbon or oxygen atom, depending on the particular class of enzyme where this reaction happened. And actually the important part of the enzymatic active site in these enzymes is oxyanine hole. It is a structural motif that is formed usually by some amino acid residues or it might be a metal cation or even sometimes it can include a water molecule. And the aim of this oxyanine hole is to polarize the carbonyl group. So from the organic chemistry textbook, we know that polarization of carbonyl group results in the accumulation of negative charge on an oxygen atom. It's delta minus and lack of the charge or lack of the electron density on the carbon atom, so called delta plus. And at this point we can say that the carbon atom is activated. It has an electrophile site and it is ready for the nucleophilic attack. So in proteins, in hydrolysis, the nucleophiles are really different. Those can be a water molecule, mainly in metalloenzymes, OH minus, usually found in binuclear metalloenzymes, and alcohol groups of serinothionine or SH group of cysteine residue. So in my talk, I'm going to cover our new results on main proteins from the SARS-CoV-2 virus. And also I'm going to talk about some examples that I found in the literature. Actually, of course, this enzyme, these main proteins is really popular stuff these days due to the pandemic. But actually we chose this protein due to its really specific features, I mean due to the interesting substrate specificity profile. So here is the active site of this enzyme. It's quite small and it is composed of cysteine residue as a nucleophile. It has cysteine residue as a proton acceptor during the reaction and actually together they form catalytic diate. The oxyanine hole is also presented and it is formed by NH groups of backbone of two amino acid residues. So speaking about the substrate specificity, it is really very specific, very pronounced. Well, here is the fragment of the substrate. So the main protein hydrolyzes peptide bonds from the polypeptides and it is obligatory for the reaction to happen or for the substrate to be recognized by the enzyme to have glutamine in the P1 position. P1 is like common name of this position for the hydrolyzes, for the proteases. It is a position prior to the cleaving C and bond. Also an important point is about P2 position. The preferred residue is lacing and actually its substitution to any other residues results in the decrease of reactivity from 2 to 50 times. So it's again about highly pronounced specificity. But up to now, there is no clear evidence of what is the origin or molecular mechanism of these substrate specificity. Of course, it is observed not only in these main proteins, but it is found in some other proteins as well. So we use molecular modeling tools, including MMM molecular dynamics and some other things that I'm going to talk later to somehow to explain the origin of substrate specificity. We suppose that the efficiency of substrate activation may be, it might be the reason of these substrate specificity observed in these main proteins. Actually, what we mean is that there are two populations of the ES enzyme substrate complexes, reactive ones colored in green and non-reactive ones colored in magenta. So if the ES complex is in green area, it can readily go to the nucleophilic attack and further chemical reaction. But if it exists in magenta area, there is no possibility for the direct nucleophilic attack. So it should first move to this green region and after that initiate chemical reaction. So the idea is to somehow evaluate the equilibrium constant between the reactive and non-reactive species in different ES complexes. I mean in complexes with different substrates and maybe this equilibrium constant can explain the origin of different substrates. So we suppose that the system can easily integrate between reactive and non-reactive species. Therefore, we perform conventional molecular dynamics simulation without any additional potentials or maybe some restrictions. But with QMM potentials and of course we treat the active side of protein of this enzyme at the QM level. The problems that we face here are the following. First of all, we should propose some criteria of assignment of conformations that we obtain along trajectory to either reactive or non-reactive. And also it's a big question of which method to use, which QM method to use for proper description of this enzyme and substrate interaction. So actually it's known from the literature that enzymatic reactions of such type are nicely described by a hybrid PV0 functional on the, well, if we speak about the potential energy surfaces. And therefore we decided to use this method, this functional as the reference one, and we take WZ2 basis set with localization functions on all ends. So next I will show you some details on, well, on the way we attribute conformations to either reactive or non-reactive. It was found many years ago, it's like paper of 1989, that electron density, that biology of electron density can be useful to analyze some chemical properties of compounds. So here is acroline molecule, and it is known that from the experimental studies that the carbon atom mark here by red asterisks is an electrophile, and it readily reacts with the nucleophiles. So, location of electron density is a sum of second derivatives of electron density over spatial coordinates. And it is known that electron density depletion regions, or in other words, electrophilic sites are characterized by positive values of laplation. And the negative values of laplation are, well, define, define electron density concentration regions, or in other words, like some structural elements like, for example, electron long pairs. So, here on the lower part of this figure we can find these electron density depletion regions in the plane, it is the map obtained in the plane orthogonal to the plane of the molecule, like it's shown here on the upper panel. Therefore, we suppose that these criteria utilized for a small organic molecule can be transferred to our complex biomolecular system. Here are two laplation maps. And actually, what we see here is that we have a substrate activation in on the left panel. Here we can see electron density depletion area or electrophilic site and bright green color in the region of sulforatum correspond to the electron long pair that is going to attack this electrophilic site. On the right panel, we can see the representative non-reactive enzyme substrate complex where we can see no activation on the carbon atom. So we decided to use this criteria, criterion as it is really easily visible and can easily be obtained from the trajectory to classify our frames. Indeed, we had quite a lot of frames, like 10,000, and therefore we decided not to calculate laplation maps at each frame, but we chose a set of frames with different sets of geometry parameters to find geometry criteria responsible for this substrate activation. So, we found that the frames with these two hydrogen bonds shorter than the selected values and also with the distance of the nucleophilic attack shorter than 3.25 angstrom are reactive. And these three criteria should be satisfied together at each frame so that the frame can be classified as a reactive. Of course, these criteria are valid only for the selected functional or the selected QMM protocol. So what we got, we calculated distributions of distances for three substrates. Well, a substrate with a laecine, the most reactive one. Also substrate with a laecine is a substrate with a second reactivity. And actually the third one was the substrate with the list pronounced reactivity with a linear P2 position. And here in this table we accumulate the fraction of reactive conformations. And we evaluated the calculated take-at-values. Here are the values normalized to the most reactive species. So we suggested that take-at-value for the substrate with laecine is unity. And then taking the ratio of the reactive conformations of other substrates and the most reactive substrates, we obtained these two values. Actually the experimental values are again normalized values and those are normalized relative to the most reactive one. We see perfect agreement that allow us to propose that this criterion that we suggested is really explaining experimental observations. After that we performed additional calculations for other substrates with other amino acid residues at P2 position. And we got a nice correlation between the calculated and experimental take-at-values. I'm not going to show these results because they're just somehow out of the topic. So then we decided to move to benchmarking and we utilized the most reactive substrate with laecine. We selected different functionals and from the hybrid function subset like PB0 with relief, the most popular one, and two range-separated functionals. Also we selected two functionals of GGA type and two basis sets with polarization functions and without them. Actually the reason was here just to try to utilize protocol that allowed to save much computational time. Here are the results of MD simulations performed at the QMM level with the GGA PB functional. And the distance distributions themselves are really close to those we saw on the histograms with PB0 functional. However, analysis of applications of electron densities at different frames resulted in the absence of reactive speeches. So that was actually surprising for us and then we tried to understand what is the origin. So we utilized a representative reactive frame from the trajectory obtained at PB0 level and we recalculated electron density with B3 leap functional and also with two GGA type functionals. In case of hybrid functional with relief, we see again this electrophilic site, whereas for both GGA functionals we see no substrate activation. Actually the main difference between hybrid and GGA functionals is that in hybrid functions there is additional term. There is a mixture of the exact Hartley-Folk exchange that is absent in the GGA functionals. And it has a non-local nature whereas if we see the formula, the GGA exchange is sort of semi-local as it is, it depends on the electron densities and the gradients at certain spatial point. Let us return to the comparison of two maps obtained with PB0 and B3 leap functionals. And actually a nice visualization of the impact of Hartley-Folk exchange here is that the electrophilic site in case of B3 leap functional is smaller than in case of PB0 that correlates with the percentage of the Hartley-Folk range. We also checked range separated functionals that have the same amount of Hartley-Folk exchange at short ranges like 20% and much more pronounced term of Hartley-Folk exchange at long ranges. So here we see really much more pronounced activation carbonyl carbon atom. So to be absolutely confident on our conclusion, we also calculated performed MD simulation with the QMM potential where QM subsystem is treated at the Hartley-Folk level. We understand all problems with no dynamic correlations on, but anyway, it's the method with the 100% Hartley-Folk exchange. And all frames from these simulations were found to be reactive. That is of course also not a good result, but anyway, we understand the impact of the Hartley-Folk exchange term. So another easily visible finger is here. It is an electron density difference map of well-obtained between electron density calculated at the PBE level and PB0 and Hartley-Folk and PB0 levels. So in case of GGA functional, we can find excess of electron density in the carbonyl carbon region shown in red lines. And contrary in case of Hartley-Folk, we should refine the sort of lack of electron density. In case of, well, we saw lack of electron density compared with the PB0 functional. So here is the last part of my results on this particular system. And what we, so we tried to use functional without polarization functions. Of course, it was not successful. We found no activation of carbon atom. Moreover, we see the very different behavior of the CO bond in these two protocols. And therefore, of course, we cannot recommend it to begin. So to sum up for this particular system, we have two findings. So one of them is more practical one. So we suggested the criterion based on laplacian electron density maps that allow us to attribute to discriminate reactive and non-reactive species. Well, complexes, ES complexes, and to use this feature to analyze reactivity of different substrates in the active site of a protein. Another important point is that the method should be really carefully selected if we deal with these complex interactions in the active site like here about this nucleophilic attack process. Next, I will show you several examples that approve my conclusion. Those are also examples of course on the enzymatic reactions starting with the nucleophilic attack. And the first one, it's really interesting stuff. It is again, of course, proteolysis reaction. And importantly here, the QAM subsystem is treated at the MP2 level. That is really not confirmed for the QMM simulations. And actually another interesting thing, another unusual thing is that the authors utilize hybrid buzzes sets. I mean, they use augmented buzzes sets with diffuse functions for the core of the active site and the rest of the active site is treated with a smaller buzzes. What they conclude is that they have two different ES complexes. The first one is formed directly after BND of the substrate. And next ES star is the one that is prepared for the further reaction. So, fortunately, supporting information included the geometry configurations of the QM subsystems and we recalculated the laplation maps from these geometry configurations. Indeed, in ES complex, there is no substrate activation and it is really non-reactive. Whereas in case of ES star complex, we definitely see substrate activation according to the location maps. So the second example is about some other hydrolase that can selectively hydrolyze caprolactone but it is inert to the caprolactone substrate. So these two substrates differ only in X atom. So in one case it is oxygen atom and the bond that is going to be cleaved is an ester bond. And in other case it is NH group and CN bond is cleaving. So according to the experimental studies, caprolactone is inert in the active side of this protein, of this enzyme. And the calculations performed in these studies demonstrate that the energy barrier is really high. It is like 27 kcal per mole. That proves from the theoretical point that the system is non-reactive. And we calculate the laplation of electron density for this structure. It is again hybrid functional here. It's PB0 with CPV double zeta buzzes set and we see no activation. So caprolactone is really easily hydrolyzed by this enzyme and it can be found also in theoretical study as the energy barrier of the bond P which is really low. And here is the location map where we can definitely see the activation of the substrate. So we can conclude that the caprolactone that is inert is not activated in the active side, therefore the reaction proceeds with extremely high energy barrier. The first example is also really nice study on the proteolysis. So it is about the CN bond cleavage peptide bond cleavage. And again we find this another type of hybrid functional BHH lip with the buzzes set with polarization functions on heavy atoms and reasonable energy barriers and nice energy profile. We utilize the coordinates of Michalis complex that is of course the same as enzyme substrate complex and calculate the laplation of electron density. So here again we see that the substrate is activated that is the reason of the vision chemical reaction that is obtained in simulation. This example is really interesting because geometry optimization is performed at the lower theory level. It is GGA type functional PBE and smaller buzzes set. And then at stationary points obtained at the PBE level, authors perform single point calculations and well with hybrid functional and triple data buzzes set. So what is interesting here, if we calculate the location of electron density obtained at the PBE level. We find no activation here. We can find this broad area of electron concentration electron density concentration. So, but if you take the same geometry but recalculate electron density at the bit really level with the triple data buzzes set, you can find the activation. The question that arises here is why do we find this high energy barrier, 27 kcal per mole if we saw if we observed substrate activation. There are two reasons. In case, well, the geometry configurations are calculated at the PBE level. And we found that this PBE complex is non-reactive. So it should lead to some unfavorable or unstable transition state. And if the stationary point is calculated to the lower level, actually there is no sense to recalculate it with a very different protocol as you just get a set of relative energies that are not any how connected to the real profile calculated at the B3 level. And the last example, it is again hybrid functional B3 leap here, and really extended buzzes set that is not that common for QMM calculations triple data buzzes set. And again we find reasonable energy barrier at the first stage, and again activation of the enzyme of the carbonyl carbon atom and enzyme substrate complex. Well, actually I decided not to put huge conclusions. As the main conclusions that meet the goals of these webinars is that QMM is really not a black box and you should carefully think of which method you are going to use to treat QM subsystem to get valuable results. And not all methods are reliable for the chemical reactions and it should be also considered when starting these simulations. Finally, I want to thank my scientific collaborators. First of all, I want to thank Professor Nymukhin and the members of the Laboratory of Chemical Cybernetics where I also work. And also I want to thank Professor Seryozon and the Department of Quantum Chemistry of Russian Chemical, Russian University of Chemical Technology, and the members of my group for from Federal Research Center of Biotechnology of Russian Academy of Sciences. So here is the last slide that remind you how to ask the questions. Thank you for your attention. Thank you very much Maria. Very interesting talk. It's very nice to see this kind of systematic approach of validation. So thank you very much. So next we will have questions and answers. Yes, I have two questions. Well, I will start with the first one. Well, as for the software that available software with the well for the elimination of electron density calculations, yes, actually, I think there are several. And what I'm using is a program called multi with an so multi like many years and do double the WFN like wave function and this program is really nice it is free and it has really many features of electron density analysis. And it is really high well rapidly developing and it is paralyzed and I really recommend it to all want to try this stuff. Okay, so I will unmute Mandar who asked this question. Yeah, so I also heard the same question that which open source code is available to calculate the Laplacian of the electrons. So you have to answer the wave wave within the multi wave within software just yeah I had the same question. And I do classical MD simulations and this is a very general question. So pardon me if it's wrong, but and I want to try QMM simulation so do you do you have any general tutorials in mind. This might not be the platform but I'm not sure whether I. I think that it depends on which software you're going to use. So now there are quite a lot of interfaces between MD classical MD programs and QM software and the most I think one of the most recent and free one is made by the team that is developing an NAMD program package for classical MD simulation and they have a set of scripts that can help to well that are responsible for the interactions between NUMD and the QM software and you can use different software's QM simulations. And I think that the tutorials are definitely the different available on the NAMD site page. Okay, thanks. Okay, go ahead with the next question if you like Laia. Yeah, next question is a bit longer. I will read it first. The beginning you said you knew from the literature that a certain functional was nice for these types of enzymatic reaction and that's why you started with this. Is there anything someone should be careful about if choosing a functional this way that you have not already mentioned regarding validation. If you didn't know from the literature what a good functional could be, how would you approach this? Well, actually there are many. I think the answer should be the following. You should think of the enzymatic reaction from the point of view of, well, general chemical reaction. I mean, of course chemical reactions in enzymes are somehow stimulated by the enzyme and so on. But anyway, they still can be attributed to the certain types of reaction. For example, here what I told is like nothing more than addition, nucleophilic addition to the activated double bond or something like Michaelis reaction and so on. And there is really plenty of literature on DFT studies of different organic reactions. Also, there are many benchmarks. I mean, benchmarks on a huge set of reactions that, like, that the claims to propose the best functional for the certain type of reactions. Maybe a couple of years ago or in maybe in 2017 there was a nice paper by a nice paper like 30 years of density functional theory. And that covers a large amount of main calculations. And I think it can be utilized as a first point, starting point when choosing functional, proper functional. I think that's it. Okay, thank you. I don't need to unmute anyone because that question was actually for me because I thought it might be useful to share that with the audience. Here is our next question that you have. What are the QMM scheme you decide decided most suitable example link atoms or boundary method. Substructive or additive. Well, actually what we use. I mean, what I'm using usually is the QMM scheme with the link atoms hydrogen link atoms and I think it is. It works nicely for these reactions where actually organic compounds are interacting. And as for the other details, we use electronic bidding scheme that is important. I didn't ask the question but it's really important. And it is, we performed some benchmarks and we showed that mechanical embedding when we do not account for the electrostatic interaction between the QM and the massive systems really give us some wrong results. It's really important to contribute this point charges to the one electron part of the Hamiltonian of PM subsystem. And what we use, I think we use it as a scheme. I mean, generally, I will unmute. How about to ask that question in order to respond. Hello everyone. I'm right on great talk. First of all, and thanks so much for answering my question. I also have a couple of questions if you don't mind me asking. So, for your work, like how big was that you am region like how many atoms for your system is specific. Well, well, it was around around 100 atoms, maybe 95 or so. So it comprised the oxyanine whole catalytic diet and a huge part of the substrate. And this makes sense. And I think in a slight 10, like, where you decided to, you know, decide on your frames. And like you have these three criteria like at three different coordinates. Well, as far as I remember you, you selected the frames that sort of satisfy these three criteria. And I'm just thinking when you are selecting this. Is it like a bit biased to select just a frame that satisfy the three criteria. Did you think about you and QMM MD simulations or like even do optimization on top of these frames. Yeah, actually, what we did we took our MD trajectory that was about 10,000 frames. And we made sort of greed. We extracted quite a lot of frames with different sets of these two distances. As of course, these, these three distances, of course, these hydrogen bonds are very important as they form oxyanine whole and the distance of nuclear field attack is important as well. And then we, well, we calculated this relation maps for a huge amount of frames. And only after that we suggested these geometry criteria and of course, after suggesting that criteria we extracted additional set of frames and check whether this criteria will fully satisfy the discrimination of frames. Right, I mean you've got the, the frames from pure MD simulation of pure classic simulations. Okay, so. No, no, no. Those are human and in the simulations. Okay, okay, so I'm sorry. Okay, okay. Okay. So just one last thing, like going back to choosing the QMM scheme, like link add-on versus boundary or this kind of stuff. Do you have like certain advices on how to benchmark this in the beginning? It seems like you did some of this benchmarking. So if I want to decide on like certain scheme of the many available schemes, what would you recommend me to look at? Well, I think that the scheme with the hydrogen link add-ons, well, it is really nicely working. It is really nicely working for these systems like enzymes and some other biological stuff. So I'm not sure that it's okay, for example, for some processes on the, well, I mean some stuff from material studies and so on, but for any chemistry, it is okay. Proteins and some chromophores in proteins and something. It is, I think the best one. Yeah, I mean, I'm sorry, you can go on. No, no, I told that in the beginning, of course, we tried different ways, but this one is like more reliable and nicely working. Yeah, it's just I have seen like some information and, you know, that quasi-project or like the chem-shell paper about like that QMM implementations. And they're sort of emphasized that it might be important not to use the link add-on scheme with like ionic systems, which might be most of the time the case with, you know, the enzymatic reaction part for like, let's say, COIP enzymes. And they recommend a boundary scheme for that. Well, here I, well, my comment is the following. Actually, when you start QMM partitioning, you should remember that some, I think some tips, like first of all, you should of course try to make this QMM border far from what you're interested in. So it's not nice if it is close to the cleaving bond or something like this. And another point is that you will better make this division between the carbon atoms with the SP3 hybridization. So it's the best way. It's not nice, for example, if you make this partitioning on the CN bond or the bond and so on. Or if you choose the carbon-carbon bond, it's usually okay. Perfect. Okay. Thanks so much. This is amazing. I really appreciate it. Thank you. So the next question is about slide five. I'm not sure, I'm not sure to have understood how you decipher between non-reactive and reactive ES complexes. So here is just our hypothesis. We propose that there might be a criteria that can help us to distinguish between non-reactive and reactive species. And then we started to think what we mean, what we ourselves mean by this reactivity. And then we came to this old paper on the reactivity and on the so reactivity in these nucleophilic addition reactions. We tried to apply it to our huge system. And if you want to go on to the next question, Maria? Yeah. How long were the simulations that you performed to access react and non-reactive conformations? Were different replicas needed? Actually, we performed simulations for 10 picoseconds. It is about the simulation time. And I didn't discuss it here in my talk, but we used a combination of NAMD, molecular dynamics software, and Terakem quantum mechanics software. And this is really fast. We used GPU and the simulation time was like one minute for one frame or one and a half minute for one frame. So you can calculate the length of how much time we did this. And as for replicas, actually, we suppose that this minima are connected by quite a low energy barrier. And we suppose that the system can easily integrate between these two minima. Therefore, we didn't use any replicas. And actually, we analyzed trajectory after simulations. And indeed, the system frequently crosses this barrier. I think I answered the question. Okay. I will unmute Rodrigo. Yeah. Thank you so much. Okay. So I guess that's the answer to the question. So feel free to go on to the next question, Maria. So are the choice of the QM region and the DFT functional and basis set correlated? Well, what I can say from my personal practice is that in case of, well, DFT functionals are, of course, preferable, much more preferable in case you want to study some chemical process with the formation and cleavage of covalent bonds. And as for the basis set, well, in QM simulations, you cannot usually use basis sets with diffused functions because they have these huge electron, well, electron density distribution. And it can artificially interact with the partial charges from the MM subsystem. And you can get some strange results. So diffused functions are not nice to be used. But of course, you can put it to some selected atoms in the very middle of the QM part, then it might be okay. And as for other things, of course, polarization functions are obligatory. And, well, usually people use double zeta basis sets as larger basis sets are really too difficult to compute. And so it's like the balance between the computational cost and accuracy. And as a result, actually, there are, I think, well, the general choice is to utilize double zeta basis set with polarization functions. And then you can just choose among several like poplite basis or CC group basis or some other. But anyway, it's not a huge choice. And moreover, we tried some benchmarks and these basis sets, I mean, the basis sets of the same size give similar results. So there is actually no that much difference. The size of the QM region is usually around 100 or 200 atoms. And this size, you can use double zeta basis set and it is more or less, well, I mean, reasonable computational time. But if you increase at least by two triple zeta basis set, it's already too difficult to look costly to calculate this. And it's not that much dependent on the size of the QM subsystem. It's just, I mean, that the the general size of the QM subsystem is like more or less determined, as I told from 100 to 200 atoms. And for this size, it's, of course, the only ways to use double zeta basis sets. So the next question is the following. If the criteria in the application of electron density, then it is dependent on the functional. If I understand, the more you there is hard to fork in the functional, the luckiest, you will find that the structure is reactive. Then how you know which function was okay? Well, so just sorry to interrupt, but this question, it might not be obvious to you based on the interface, but this question follows on from the earlier question from Isabel, which said, which referred to the question started with, I'm not sure to have understood how you deciphered between NES and RES, which you answered. And then she responded by saying, thank you. And therefore, then this this follow on came. So the question in the ending of this comment is that how you know which function is okay? Actually, of course, we cannot be 100% confident of on the functional, but we know we have quite a lot of literature of the already published results. And we know which functions are more or less okay for the certain types of reactions. And also, I think that if we obtain this, well, nice, nice correlation between the experiment and our calculation, and as it was also brought into other substrates and the correlation still exists, I think it's the indication that our method is okay. I mean that I think that it's not only about this case, but you can say that, well, your method, your simulation method is adequate if you get results that can explain experimental data. It's one of the indications that your method is okay. I don't have any more questions. Okay. So in that case, thank you again very much. On behalf of everybody who attended and who cannot literally clap with you being able to hear so. Thank you very much for the presentation. As I said, I thought it was a very nice example of kind of systematic validation. So the next webinar that we have will be in a couple of weeks time. And Ulfrüde from Lund University will be talking about obtaining accurate structures and energies using QMM. The link, it's on the BioExcel website. You can register now if you like. And there's a short link there. You can always follow us on Twitter as well to stay aware of the latest updates. So yeah, the link there is to the overall event page for this workshop as a whole. So hope you find this useful and look forward to seeing some of you again at the remaining webinars and ultimate panel discussion for the workshop. And Maria, thank you again very much for your presentation. Okay. Bye-bye everyone.