 Okay, there is a question regarding the Born-Oppenheimer approximation in QM calculation. How can we compute the nuclear contribution for energy refinement? Usually in Gaussian software we perform a calculation to obtain the thermal corrections. Okay, here it's not clear to me the question because at least I have in mind two different things and so I will try to clarify a bit what I have in mind in this moment. If you talk about thermal corrections in the sense like the corrections to the Gibbs free energy, if I remember well as I said are called or the correction to the enthalpy that are usually obtained from Gaussian with the thermochemistry calculation kind of calculations, then I suppose that I know that you can get the same kind of quantities also with CP2K using doing the standard vibrational analysis calculation, if probably you have to introduce the thermochemistry keywords if I remember well, but maybe here Dimitri knows better than me. However, I want to point out that this kind of approach that I was referring to the thermochemistry approach is valid for molecule in molecule in gas phase. So you do single point calculation and in addition you have to say suppose make the approximation that all the particles are not interacting. So your system is an idea I guess and far more far far as you are from this approximation and less accurate will be your results. Quite often if you have a code that allows you to do molecular dynamics, you can instead adopt a different approach of course computational more expensive because we have to do a molecular dynamics. This is the typical approach that user of CP2K is PMD quantum express or users etc. usually adopt which is not to do single point calculation plus the assumption of the adiagas but sampling directly a statistical ensemble via an ab initio simulation. And this is just a thing that I had in mind. However, in the first part of the question is regarding the Borno-Pena approximation in quantum mechanics, how can we compute the nuclear contribution for energy refinement? Okay, from the first part I started to remind another kind of topic argument here in the sense with the Borno-Pena approximation let us say you and in general with ab initio molecular dynamics schemes you decide to treat the nuclei or the core electrons in the core part of your atom not only just as a let us say classical object which are described by Newtonian like equations, can we go beyond this approximation? The answer is that in general yes, this means that we should start treating quantum mechanically also the nuclei. These are in general called nuclear quantum effects is a topic not so popular because the nuclear quantum effects are relevant in very specific area or context in chemistry not in general are totally negligible but in principle this is possible. One approach that I have in mind is called ab initio part integral molecular dynamics. I'm not sure that this is implemented in CP2K for sure I know that was implemented in CPMD but of course this is only just a keyword to give you to you in case some money in the future would be interested to this nuclear quantum effects and I was in my mind came out these things just reading the first part of that question. So with regards to that, thank you Miliano. I saw also another question that may be interesting. Can we generate metal ion parameters from the QMM setup to then only using MM? First of all, in principle yes, first of all I would remind the observation that Dimitri did this morning regarding the fact that when you parametrize, you want to parametrize some objects, some ligands, some residue because you do not have parameter with the aim to do maybe a very long classical molecular dynamics ideally if possible you should always look at how the other parameter of your forces that you are going to use were parametrized in particular if they were parametrized through high level quantum chemistry calculations and follow as much as possible the same approach that the developer of that force will use this because you would like that your new force field, no parameter let us say is totally compatible with the rest of parameters that are in your force field and that you will use in the simulation, in that specific simulation it is true that sometimes this is not possible or at least could be is questionable and in particular for example with metals and so in those cases of course it is possible to use for example CP2K to let us say create and obtain a DOC parameters the keyword here is force matching you should find a section regarding force matching in the city's way manual let us say to see some details what I can say here of course one thing that you have always in mind is that when you get a parameter or some parameter of the force in from this in this way this parameter probably will work very well for your classical molecular drama simulation if you obtain them from a QMM simulation coming from CP2K for example however this parameter are absolutely very probable not transferable in the sense that this parameter are very suitable for that specific system from which you have used to got them but cannot be transferred for other system either even with the rather similar let us say chemical environment for example around the metal ion or at least you could not be confident that it could use easily let us say or successfully with another system or even very similar you support the question about which question you are answering so there is an answer for the moment the third one first one about the destruction of vibrational frequency calculation now it is not possible with the interface unfortunately because grommets have no such functionality and vibrational frequencies also vibrational frequency calculation with QMM is kind of very questionable because the system is too big I know that there is some kind of approaches to do that but I personally never trust it it is much better and if you look into our way we are serious about the best practice in QMM you will see that most of the people now instead of doing this transition state structure they are more moving towards the free energy profile of your reaction using QMM molecular dynamics so kinetics is much better reproduced with dynamics so to say so it is always better to do umbrella sampling for example because umbrella sampling will give you the same result as will give you better results than your transition state structure because your transition state structure is very dependent on the your starting configuration MD is usually not fully dependent so what is the best way to refine energy at high theoretical level in order to obtain more accurate results, activation, deep free energies here I can forward you to our parcel youtube channel to our best practice webinar series in the QMM which I will have a link in my document in the presentation in the last slide here there was several examples what you should do and how you can do that for example how you need to check the size of your QMM system for example is your final results will be pretty much dependent on how big is your QMM system and how you can check that so this is a very long answer by itself so another question is looking, do you have something else to add here? to me today I was going to say meanwhile the chat you were having is there any of the questions that you want to say anything else about I know you responded to a few different ones yes there was a question about link atoms which appear from the fighter counter apology if you look into that but I think it is old one they just sent it out they are doing nothing yeah I just forgot to throw them out link atoms are now in the interface generated automatically so there was also a question about how you should really calculate the spectra so really if you want to calculate the spectra what I suggest to do is to take a long dynamical trajectory that could sample the conformational space of your protein well then you need to clusterize that trajectory somehow take a significant clusters from it take a snapshot from that cluster and start a QMM trajectory from that so the idea is that with classical MD you can sample on a long time scale with very cheap method and find the global kind of conformations which protein could get and with QMM you can access the properties you want on a shorter time scale yes Arno also have a link to that best practice workshop and you took the list so which you can also look into and there is probably the answer to your questions how and what how you could do this yeah Emiliano have you something to add on that I saw simply the last question in the in the document is about could you comment on the usage of semi-empirical methods with the GROMAC CP2K interface are they as easy as DFT calculations maybe I suppose that in the interface you can do semi-empirical or DFT without any problem am I right? yes but for now semi-empirical you need to use your old CP2K input like you've done now this so it will set up themselves how to deal with the dangling bonds now it is treated automatically it will put link outcomes automatically on the dangling bond but what you need to be sure is that you're cutting the right bond so for example it is not a very good idea to cut over the pi system and have a very strong quantum interactions and if you cut for example your benzene ring into two halves it will not work so you need the ideal situation where you need to cut you need to cut over the sigma bond and more importantly that sigma bond is the best one you can cut should be between the two carbon atoms in my experience so for example if you cut between C-alpha and C-beta you need to cut over the bonds in the protein chain it will be okay also it is very good where you can cut if you want to cut with a backbone of your amino acid then it is better to cut over the C-alpha bonds so CO and C-alpha bonds so this bond needs to be cut also you should not cut in the protein you should not cut over the peptide bond in the pi system nitrogen, carbon and oxygen form the peptide system forms the pi system so you should not cut that cut on the C-alpha bond this is the best way how you can cut the protein by bone and of course never cut on the C-C bond in a ring or something like that in the pi system because the idea is that never cut the pi system in a bioalpha system it is a really bad idea new big question, do CP2K has some GPU speed up I think Arna Kud commented on that I think we discussed it earlier today already so CP2K certainly does support GPU it has CUDA based offloading for NVIDIA GPUs natively within the application therefore it's via a library that it uses for sparse matrix multiplication which is used for some of the function calls within CP2K it can also do some offloading to AMD GPUs but that is of course a subset, it's not very mature it's also a subset of the offloading that in principle is possible so it definitely has CUDA offloading for GPUs which is being developed quite actively at the moment whether it actually gives you an advantage depends on your system a bit CP2K has, broadly speaking, two performance regimes depending a little bit on the size of your system and the size of your basis set CP2K is often quoted as linear scaling code and it can scale very well to very large systems but that's when it uses slightly different approaches which also makes more use of this sparse it's also about the sparse of your system so that also makes more use of the sparse matrix multiplication library that I mentioned which therefore has more additional offloading so it's difficult to say in any case in advance whether it will benefit from GPU the question is you should try it for a lot of systems traditionally it's been the case that it's not necessarily for the case of having say compute nodes where you've got maybe two Intel Xeon processors, maybe 24-36 cores with some generation of NVIDIA like Pascal P100 or even maybe the Vultas there's not necessarily that often an advantage of using the GPUs additionally for offloading compared to use running purely on CPU but the point is you have to try it trying the usage with Gromax plus CP2K I think that already came up earlier today we're looking at characterizing the performance providing guidance on this so we will update the best practice guide which we refer to today which is the QMM best practice guide as Holly said for standalone CP2K we will update that and expand it with guidance on using Gromax together with CP2K to run on GPUs as I mentioned earlier you can use we know Gromax works very well on GPUs but the Gromax execution of the classical part is of course not the bottleneck limiting factor in the calculations here shall we look at some of the other ideas? Yeah, what is the best way to refine the energy high-tech for I think also answer to that the best idea is to sort of look at webinar series on best practice in QMM there was a lot of talks about this accurate reaction and activation energy also with validation of your size of QMM system for example and higher level methods I think who is Ulf Ruder he talked about this higher level of theory and validation of QMM models Yes, Arne? Yeah, I think Ulf Ruder was talking about looking at systematically increasing the region size and number of QMM and having guidelines so his webinar talks about guidelines rules of thumb for how far away from the active site and how you know the effective inclusion of more and more I think Adrian Moholland's talk also talked about different levels of theory in it Okay, so looking at how two small molecules react with one another through QMM I did not quite question what do you mean if you mean reacting molecules, reacting molecules should react with QMM because in MMM you don't have any on breaking typically so yeah, I would suggest to clarify the question Yes, you can look into small molecules reacting with one another using for example the same umbrella sampling technique no problem critical perspectives on dealing with non-parameterized ligand modified residues can trust under chamber I think it's a more general question about the molecular mechanics rather than QMM but what I can add here is that if you will follow the anti-chamber protocol what it uses to parametrize your pamper is a 14-amber 0.3 force field yeah, you could trust I mean the whole amber force field is built upon the same approximation so if you do it quite accurately then yes, at least it will be not less you do not produce patch error in comparison to the other residues so what are some pitfalls what we could make our own decisions there is a whole discussion about the wrong decisions again in the best practice workshop so if you're really interested in that please follow that I think in the final panel discussion there was talked a lot about the wrong decisions there are failures so even the high I mean it's nothing wrong to be wrong yeah so a lot of scientists even a high level of scientists who are good at workshops that they also show where they are failed and sometimes they fail really critically and they need to scratch the years of job years of computations or reduce them and there is nothing wrong in just science there is no success without how it's called, without downfall I would say that one of the take home messages from that best practice workshop part of the reason we organized it was that it's I mean it's great that we have the interface with Gromax to CP2K it's just the only thing to be aware of if you're using this it's just that it don't not to treat it really like a black box to be aware of that but I would like to move through this course of the relative benefits or pros and cons of the different functionals what is included in different levels of theory and not to just blindly use a functional because well it's the default and you know it's probably fine it might be but try to look at looking at validating this somehow and the best practice workshop gives some examples of how you can know about webinar talks about one example of how you can validate things so this is something to be aware of and that's about the QMM simulation approach as a whole or modeling approach rather than using CP2K for example I mean it's just true whatever software you use yeah it's always will be true for any software you use you always need to validate what you are doing also in classical molecular dynamic solution you also need to do for example see that your system starts deviating a lot from the crystal structure it may be already a warning that you are doing something wrong usually but yeah additional tutorials practicing QMM simulations with different scenarios using proteins after the course can you generate metal and parameters for QMM and only using MMM okay so it's again about the parameter generation for parameter generation it's better to follow the protocol and generate how usually they are generated using the using the quantum mechanics indeed in the course but you need to follow the protocol pretty closely yeah for under chamber or for char GUI I guess also that's the second one who can produce you so yes so there's an idea can you extract the parameters on QMM but it's not the best idea to do that because usually you are forced to this parameterize it is a different level of theory okay can you go a bit further how can we use chromax Cp2k to calculate reaction profile mainly it's a heavy abstraction so there is now one way which I showed to you is umbrella sampling you can do umbrella sampling and you can perform this kind of standard approach with most of the people using in real life in real situations because it's the most convenient way for now we're also looking at some additional ways how you can do this for example if you look into the Microsoft there is now method called accelerated weighted histogram free energy method which supposed to be a bit better it should converge faster but for now it is still under the consideration if there will be such things I think I will make an example of how we can use that it's one of the targets for us now okay there is some question in the chat also about the DFTB if you want to do DFTB with the interface technically Cp2k have DFTB capabilities and you can do that but it's on your own risk because Cp2k Cp2k implementation of DFTB is not quite well in writing and it's all very stable and many people complains about that so for now I would suggest you still to use all the implementation of Thomas Kubart with a 28 in chromax I guess but we are looking into involving also Thomas maybe somehow and make the DFTB interface with a newer version of chromax it's one of our next targets probably so about the machine learning approaches so Andrew asking so they are playing now with ASIA-ANI force field I guess that machine and they want how it can be combined also with that with the chromax so yes indeed we are not done on the interface but also our colleagues from TTH they've done a whole application programming interface how you can implement your own force fields any external forces and so on by yourself yeah if you are skilled enough in the C++ programming and one of the examples how this interface works is our Cp2k interface it also works through that there is also I think density fitting approach it works through that API down in chromax so yeah there is a whole machinery using which you can implement your own force providers so you can provide any external force and chromax will integrate something to the simulation so yeah it looks very promising I also promised at some point to do this to maybe implement this ASIA-ANI force field into the chromax using this but we'll see so I think it's worth to try at least extension to the reaction profile question if one needs to break a bond the atom transfer from ligand to protein can that just be done with a pull command in a Braille sampling or there is another way the best idea how you can do this reaction profile how I suggest you to do that reaction profile in general with a Braille sampling so first you do pulling simulation where you pull your system from state A to state B using this pulling with non-zero velocity so you basically put this pulling velocity to some value and pull the system from one state to another it will create you some structures along the pathway then you pick up that structure evenly spaced along your reaction coordinate and start a Braille simulation so that with a zero velocity that's usually how it is done now in the Braille simulations so it needs to be done through the pulling coordinate this answer to your question the way how you do this in Gromax is pull it's pull code pulling code so that is okay so for metal lines so it continues also for metal lines yeah I think Andrew gives really good paper on parameterization of metal lines also don't forget that the metal lines probably if you want to parameterize them well you need also to parameterize them in complex with the amino acids which they are bound towards yeah so you need to make one huge residue which will contain metal ion and amino acids which are connected to it yeah so this is how you should do that in reality so I think we covered all the questions so which I saw in the chat in the file also also about the metal ions I would suggest you really once you do in the metal ions with CP2K for example please be very careful with the basis that you are using and with cutoff for the plane lights you are using so metal ions is one of the things which are kind of not so easy to converge because they have usually a very high speed it could have for example iron ions they could have a high spin states higher spin states in the singlet and this will really slows down simulations and lead to some numerical instability in quantum mechanics so always check for if you have a metal ion in your system and you can use it to do the measurements of cutoff and take a big basis that's what I suggest you to do okay well I can just wrap up then if there are no more questions again we've talked about it a couple of times already we've raised it again and again because hopefully this will be useful so if you want to continue today or tomorrow to sort of extend you can just basically attend this workshop retrospectively by looking at the listening to the webinars and the panel discussion so here lead here some of these external speakers, the experts together with Gerhard Ruhnoff who is also in BioXcel tried to tackle some of these real challenges surrounding the use of QMM for modeling simulation much more generally so not specific to CP2K so a lot of the aspects that might be useful there to look at how people actually tackle these issues what things can go wrong, what to look out for so when it comes to using Gromax with CP2K or CP2K standalone we have BioXcel has a support form ask.bioxcel.eu there is a section on the form for QMM so actually in the schedule on the course materials page in the wrap up section that I've put in just recently there's a whole list of links there's a link to the form there as well so finally I just wanted to say well let me quickly just share the page that I'm talking about just to point out some of the links there so yeah so we have here the slides the feedback server links there then there's a link to the Dimitri's modified version of Gromax with the interface integrated and there's a lot of different branches the branch is relevant for the version of Gromax used on R22 during the course is a 2021.1 Plumat branch installation instructions are in install dev as Dimitri mentioned there are links there to the QMM best practice guide to the workshop there's a link to a paper there which might be useful when it comes to considering functional it's sort of a warning a guide and a warning to the zoo of available DFT functional so that might be something to help you make some choices as well. A link to ask by Excel form and there's a link to various kind of useful maybe scripts and further tutorials developed at least in part by us that might help you in future so with that I would like to thank Dimitri and Emiliano for contributions Holly for her session as well and I'd like to all thank you all very much for attending and do get in touch on the forum if you have any other questions to follow up on other than that I wish you very good rest of your day Yep, very good. Thank you Thank you for your link whatever you have