 Welcome, everybody. I was just waiting that everybody has the chance to get in. And we see, it seems that we have a stable number of attendees now. Okay, so welcome to the BIPESAL webinar number 76. Today's presenter is Banjo Lunbori, he's from KTH, the Royal Institute of Technology Stockholm, and he will speak about the GROMACS 2024, and in particular the new feature improvement. I'm Alessandra Villa, I'm together with Otto Andersson, I'm hosting this webinar for BIPESAL. So during the webinar is recorded, so just that you are aware of that. And during the webinar anytime, you can use the Q&A function to ask questions. So you will see this symbol, or this one, at the bottom of the Zoom application depends on which operating system are you using. And then you can click and type your question, only not everybody will see your question, but we will see it. So at the end of the webinar, we will unmute you if you have a microphone. And so you can ask directly the question to Manus, otherwise I will read the question for you. So it will be useful if you also give me the information if you don't have a microphone or if you don't want to speak any way. Or you cannot speak, that is also another option. Because I cannot see it. After the webinar, if you still have some questions, in particular maybe if we have a lot of questions that are not answered by Manus, you are most welcome to ask them in the forum, and Gromax has a dedicated forum under Gromax.byexcel.eu, and in particularly under the category user discussion. So something about Manus. Manus is a researcher in Eric Lindell's group, and is currently leading the Gromax development. He's involved in Gromax development since 2012, and that moment is just start as a postdoc in Eric Lindell's lab. In the period between 2015 and 2023, so just last year, he was working for a company, Ergo Pharma, and he was particularly involved in calculating skin permeability. He was in molecular dynamics simulation. He has a master's in pharmacy from Uppsala University, and he got his PhD in organic chemistry at Stockholm University at the end of 2011. After that, he was a postdoc for one year at the biochemistry department at University of Cambridge. And then I think he moved to Stockholm again, and now he's currently working in Stockholm. Okay, so now we will see what he tells us about Gromax 2024. So I stop sharing, and then Manus can share. So, welcome to this webinar. And as Alessandra said, I will present new features in Gromax 2024, but I will also spend just a few minutes going through or recapping what happened in the 2023 development. So first, starting, what is Gromax? I think most of you already know this, but in case we have any listeners or viewers who haven't used Gromax yet, it is a molecular dynamics simulation package. So you could consider it a virtual microscope, where you can follow atomistic movement over time in detail. And we nowadays say that Gromax is user driven regarding the development that we try to follow what users need. And we try to implement that to the best of our ability and our resources. Gromax is considered performing very well in that it has a high speed. And it scales well to large computers. You can use CPUs and GPUs in combination. And it runs on most platforms and hardware. So it's very flexible in that way. It also includes a suite of analysis tools that you can use to analyze your simulation results. And we also provide an extensive documentation and also then tutorials to help users use both the simulation options but also the analysis tools. As I will go through a little bit more in detail today, there are also interfaces to other software. For example, to CP2K for QMM simulations and also the new Colovars integration that was added in 2024. Gromax is free software so you can use it without paying and you can modify it and you can redistribute it and do more or less whatever you want with it. As long as you acknowledge the original work. And these are some links that may be useful. So just a brief going through where you can find information about the software packages and its documentation and all. So as Alessandra mentioned where the forums are located, if you are interested in the development or if there are any issues that you would like to report, then follow the last link here, the development link. So Gromax 2024 had its first release a little bit more than a month ago. And we have these yearly releases of Gromax. And then the first patch release where we fixed some discovered issues was released last week. And we expect to have a next patch release in approximately two months. So what happened in this last release in the 2024 release so one of the main features is the integration of Colovars into Gromax. We also improved the accelerated weight histogram method, which is an enhanced sampling method with the possibility to automatically scale the target distribution. We also had improvements to the deform option and now you can calculate shear viscosity is using it. There are also a few things that were addressed. For example, now the artifacts that were present from Missingland Jones parent directions during the course of an interaction list, where the lifetime of an interaction list. Now that has been reduced. And there were also performance improvements and I will go through these features. A little bit, but I will focus more on the Colvars and AWH improvements. And as I said, the Colvars integration now means that you can reuse Colvars easily from Gromax just with the MDP settings and a separate settings file. And Colvars enables the use of quite advanced or sophisticated collective variables for your sampling. And you can then apply restraints or biases to these. And that can be used also then for enhanced sampling, for example, using the adaptive biasing force method, which I will go through a little bit here. And just keep in mind that I'm just presenting Colvars here as an example. There will be a new or in the Colvars webinar in two weeks by the Colvars developers, which will go into more detail there. So this is just sort of from my perspective, my first impression or examples of what you can do. So to illustrate this, I will use an example that was published by Noraho and others from Chemarchive. So I have used their input data, so their coordinates and their MDP settings. And I have used these examples both for illustrating how to use Colvars and also as examples to the improvements to the accelerated weight histogram method. So in this case, we are looking at the binding or unbinding of bensamadine to trypsine. And it's in principle a fairly trivial setup in that much of the protein backbone is restrained and also the ligand is all only pulled in one dimension and this part from that largely restrained. In theory, it would be an efficient system for predicting the PMF or the potential of mean force as you pull the ligand away from the binding protein. And to set up these Colvars simulations, there are two options in the MDP file that I have been using. So you just say that you want to use Colvars and then you specify the name of the Colvars settings file. And in that file, you then specify the settings that you would like to use for your Colvars, in this case, enhanced sampling simulation. And these settings should be taken with a grain of salt because I'm not a Colvars expert. I just looked at the Colvars documentation and tried to make a rough translation of the AWH settings that were used in the binding study that I started from and then applied that in Colvars. So you specify your collective variable, you specify the range you want to study and what kind of collective variable, in this case, a one dimensional distance between two index groups. And then you say that you use the adaptive biasing force method for the for enhanced sampling. And I also specified upper and lower boundaries for my sampling here to prevent the ligand from too far away from the range that I'm interested in. And then you get a PMF like this. And this is from a hundred and a second simulation of only one walker. So this is just an example of the output that you get. And I will then in a moment compare or show output from AWH, but that doesn't mean that we should compare these because this is only from one simulation. So it's just to show that this is the output I get from a very simple setup and the performance is comparable to AWH as well. But looking at this, one thing that might be interested in is to see how much did I sample in different parts of this collective variable or the region that I've been pulling the ligand from the protein. And you can directly get that from Colvars as well. So you can just in one of the output files, you can get the count in different parts of the sampling range. And in this case, we can see that the count is very unevenly distributed. I presume that if I had simulated longer, it would be more even. But still you see that in the higher values here where you are mostly having the benzamidine in a water environment, we sample a lot more. Whereas in the low distance ranges here, we sample very little. And now I will use this as an example how to couple this to the accelerated weight histogram method. So from the AWH method in the output, you can get an estimation of the friction metric along the free angle landscape. And that would look using the same Colvars sampling. And then in red, we have the inverted friction matrix. You could actually call it the diffusion metric from AWH. And you see that it's very noisy, but we also see that in regions where we have a high calculated diffusion from AWH, we also have lots of sampling or more sampling from Colvars in general. Even if it doesn't follow exactly. If we would then instead print or plot the output of the AWH friction metric. So instead of taking the inverted friction metric, it would look like this. And this is one of the new features in AWH that you can use this friction metric to scale the sampling or to scale the target distribution. So that you steer the sampling to regions where the friction is high or the diffusion is low. So the regions that would be slow to sample, get higher bias to sample them more. And this is enabled as an MDP option that you just turn it on and say that, yes, I want to scale the metric or the target distribution based on the friction metric from AWH. And then of course the results of this will be very system dependent. But looking at the same simulations that we did before, just plotted here more extensive simulations than I had from Colvars. So here I have the output from five repeats using four walkers. So from the Norahu setup. And then I did the same using this scaling of the target distribution. And what we see here is that the standard deviation goes down in this case by factor almost two and a half. Which means that we have improved the sampling efficiency by a factor two and a half four or five to six approximately. So by sampling more in the region where where the binding takes place or whether there is tight interaction, we have now improved the sampling efficiency significantly. This is just one example but using the same setup still but then plotting it with them. In this case I have four plots that blue and the orange are using one walker. Four repeats and the red and the green are the same as I had in the last plot. And here we see the same trend that since that I have fewer repeats with the first in the blue and orange. I don't there is no point in comparing them to the one with four walkers. But at least we can see the same trend that the standard deviation has now here gone down by factor three or the one walker simulations, which in turn means approximately a speed up of a factor of nine. And we have seen these improvements from between a factor of two to a factor of 10 in differences simulation setups, but there have also been cases where there is no improvement at all. So it depends very much on the friction metric across the simulation or free energy landscape. There is one more option in a WH that was added as well. And that is the growth factor during the initial stage. So how quickly the histograms grow that previously had a default value or hard coded value actually that was set to three. But from 2024 now we can set it in the mdp file and the new default value is set to two. But if you use old TPR files, it will use the factor three that was used before. So the factor two means that the growth in the initial stage is a little bit more. Well, careful, it doesn't grow quite as quickly anymore. Which means that it might take a little bit longer to leave the initial stage. But on the other hand, there is a little bit less risk that you have sampled it too quickly and perhaps got a little bit too high free energy barriers in parts of your free energy landscape. There is also now, or there have been some improvements also to the deform option, which means that you can use the deform option for calculating clear viscosity property now. So you deform the simulation box. And you generate the velocity profile that is applied during the simulation. And this velocity profile or the flow field is generated at the beginning of the simulation. And I will just show a few examples of how this can be used. So first I will just illustrate the actual deform option by pulling a system or deforming a system in the z direction. And then I just use this in it flows to set the velocity profile. So this is an example of a lipid barrier system. So from the only layer in skin or the stratum corneum. Where we have a very tight lipid barrier. And what I have done here is just that I have deformed the option in the box in the z direction. And just looking at this, we only see that the box is extended, but we don't. It's a little bit difficult at first sight to see what effect this has. But if we show the periodic images of the box, we see that we have pulled the system apart in the z direction. And that is quite natural that you will do after a while when you scale the box like this. What is a little bit interesting to see here is that as you might expect, you pull it apart at the point or the interface where there is less tight interaction between the lipids. Whereas in the other interface in the middle of the barrier system, there is more interaction between the lipids. They are interdigitated and there the connection is still retained. But this was just an example and doesn't really help you doing much new science. To illustrate the actual more interesting deformed occurrence, I have just shared the system using this MDP setup. And then using the same system as before, if I share it, you first now see that the simulation box is tilted. And you also see then that the chains are getting correspondingly tilted as well, or at least the tilt is no longer the same as it was before. So they are getting more straightened out where they were tilted in the other direction before. So you see that the system is getting shared here. And using examples then from Michele Pellegrino, who has been involved in this development. We can see that we can use it for calculating the shear viscosity. So here it's just a box of liquid that is sheared. And he has been using this for calculating the shear viscosity in different solvents and solvent mixtures. And what is interesting to see here is that using this method, it gets a very small or in general very small error bars from 100 nanoseconds of simulation. And he calculates the shear viscosity using the Einstein formula now in the GMX energy tool. And his development has also improved the efficiency of that shear viscosity calculation to make it quicker. As I said, other improvements that we have seen or one or the other is that previously there were reports of pressure drift during interaction pair lists. And now you can set how large you allow this pressure drift to be. And this bug or issue only affected systems without Coulomb interactions. So for example, some coarse grain systems would be affected by this. And with this reduction of the artifacts, so the new improvements in 2024, there might be a slight performance loss for these coarse grain systems in order to avoid this pressure drift. So you might see a little bit of slowdown here if you're running coarse grain systems. To recap a little bit on the performance improvements that have happened already during the in the 2023 release, we had major improvements or in the sickle GPU features. So now sickle is working on most or as far as I know all important HPC platforms. One example is Lumie in Finland, where we use or where sickle is used by default. Another feature that was introduced in 2023 was the put a graphs and that means that you can record the GPU activities during one simulation step. And then you can reuse the same sort of mapping during the following steps that you already know exactly how the calculations will be done so that you don't have to launch the kernels in the same way. So it's more efficient in that way. This was enabled using a separate environment variable. I must admit, I don't know if it's enabled in any other way in 2024, I think it's still using the environment variable. So it's a little bit experimental still. But it can speed up your calculations if you are using CUDA GPUs. You can also decompose PME or divide PME calculations over multiple GPUs nowadays. And also updates and constraints are run on GPU by default since 2023. And that of course only applies as long as your simulation settings are compatible with running updates on GPU. There were also some modifications on the default coupling intervals on thermostats and borostats, which has improved performance slightly. And for 2024, there have also been performance improvements. There have been continued circle improvements. But of the new features, what you might see is that you can now use the mass repartitioning function from grow MPP. So that when you run your simulations, you can always scale your hydrogen masses and shift that mass to its connecting heavy atom. And this means that you can, since that will affect the bond vibrations, you can now, in most cases, if you set the mass repartition factor to three, you can often scale your time steps by a factor of two. But this is system dependent and it will sometimes not make that large improvement. This has previously been available in PDB to GMX. But especially if you are using systems where you might, well, if you have, for example, ligands or other groups that are not treated by PDB to GMX, now you can instead just do all of it from grow MPP instead, which simplifies things. There are also in the 2024 release, there are more options for how to use the multiple GPUs for PME calculations. So you can now specify those settings using environment variables, which means that you can often tailor this better to your hardware. And I assume also that you can tailor it to your specific systems a little bit more efficiently. And importantly, I want to thank everyone who has been involved in the 2024 release. And of course, this would not have been possible without all developers involved in previous releases either. But these are the ones that have been involved in this one. So thanks to all of these people, it would also not have been possible without funding. And these are the main funding sources or external contributors. So we have dedicated development from NVIDIA or sorry, we have dedicated development from Intel and we have developers at NVIDIA who have been helping us significantly, but they are not dedicated to the grow MPP project. And more recently, we also have a developer from AMD who is working for us, but that did not make or did not fund the development in the 2024 release, but we will see that in upcoming releases. And that is where I will end this and leave this to the questions and answers. I hope you have lots of questions regarding this and for new features. Thank you very much. Thank you. So we have, please, I will start with question from and we have a question from Pedro. So I will try to unmute Pedro. Just give me a moment. Oh, I make. Okay. Okay. Yes. Can you hear me? Yes. Yes, Peter, please. Yes. The question is cold bars in your implementation faster than, for instance, plummet. I haven't compared them. So I don't know. I would think that it is probably faster, but I won't promise anything. Going to another question is, is a ABF method supported in Gromax? And not without cold bars. Okay. But with the cold bars, it's, as I showed here, it was very easy for me to set up and I hadn't used it before. And I presume that there will be more information about the adaptive biasing force or ABF method and also many other cold bars features in the cold bars webinar in two weeks. Yeah. Okay. My last question is, if there is any update on the fast multiple methods that is being developed by by a group in Germany? I would say both yes and no. So there is ongoing development of the fast multiple method. Also, well, as you say, there is one group in Germany doing development. And there is also development from the core Gromax team in collaboration with one group in Japan. So there are two separate implementations that are being developed. And at least the plan is to include the variant or alternative that has been developed together with the group in Japan and hopefully in the 2025 release. That's the plan, at least. Thank you. Thank you, Pedro. So now we go to Johan. Johan, I don't know if I'm sorry if I pronounced wrong. I will allow you to talk. Okay. Can everybody hear me? Yes, please go ahead. Cool. So yeah, I wanted to ask if it's possible to develop our own CVs with cold bars. And the reason I'm asking that, and this may answer Pedro's first question, is that using plume, you need atom coordinates on the CPU. So if you use plume with Gromax, you've got to run updates and constraints on the CPU. So obviously it's, it's lower. So is it possible and how easy or straightforward to implement our own custom CVs using cold bars? I must say that the cold bars developers would probably be better at answering this, but at least my impression is that cold bars gives you a very flexible interface for designing and setting up your collective variables. So knowing your exact needs, I cannot say if it will cover exactly what you are doing in plume or plume D. But I think if you have a look at the cold bars manual, you will hopefully have good examples there. Thank you. Thank you. Thank you very much. So now there is a Lila. I will allow you to talk. Please go ahead. Hello, can you hear me? Yes, please go ahead. So I have a quick question about the deform part. Can you please talk a little bit more about that area? And then is there any tutorial that we can follow and learn? Or is there any template MDB file for the deform option? So in this case, there is no, as far as I know, at least no tutorial. So the deform option is, it has been around for a while, but it has been now improved that you don't actually move the atoms when you deform the system, but you instead apply or use a velocity profile. But for example, this is the only input I or the only deform relevant input I used to share the system like this. So like, if I want to repeat the same thing, I'll use the same MDB file and add these two lines so I can somehow make the next slides you made, right? Yes, exactly. And what you might need is to adjust this 0.001 value to make it deform as in the speed that you want. And my last question is, is there any limitations for the speed? Like here you're putting 0.001, like is there any limitations on the range that we can put here? I think it will depend on your system. If you hear it very fast, I think you might see artifacts. So I would say that now I know the exact numbers, but I think this is in nanometers per picosecond. Yes. So already these values are quite high, but I just used them for to generate quick examples here to illustrate what the deform system looks like. I mainly included this example because the pure fluid examples make it very difficult to see what has actually happened. If we just look at this picture, we can't really see that much has happened to the simulation box except that the periodic box has been shared, but we don't see any movements of the liquid as such. Understood. Thank you so much. Yeah, thank you. Okay, so I have one other question to Laila. So I want just to point out one, will come a tutorial. I don't can tell you exactly when, but it will be a tutorial on the formation of a gram max. I will guess not, maybe in a couple of months, I hope, just for your information. Thank you. Thank you. So now the next speaker is. And show again, but before I want to just to point that please the call bar question. If you could put in the forum. So I just have online that is also following the webinar is also fioring that he will speak in two weeks and he asked if kindly, you can put in the next forum I will put the link to the gram max forum in in the chat. And then if you can type there your question so people call back people can answer there. Thank you very much. Thank you very much for pointing out. So then we go further. I will say next week. Next, I tried one note, sorry that I have to find it. Sure young one, please, if you can you can speak. No, no, I think you cannot. Okay. For my understanding remarks are not currently sport to pure acceleration for stochastic dynamic integrator. Is there any plan or likelihood that this capability will be introduced in the future. Yes, there is a plan to introduce that in the 2025 versions there is an ongoing effort and if you are very experimental it might actually be possible to test it but I would recommend to wait a little bit. But the plan is definitely to have it in 2025 releases. Okay. Yeah, yeah, you know, because that's very important for the free energy calculation. Yeah. Yes, exactly. I know I've been seen personally as well. And then last question is, you know, now they grow max can use a plume to perform lots of the acceleration enhanced sampling like meta dynamics. So what's the biggest difference between the plume and colvars? Yeah, I should leave that to the plume and colvars experts but I would say that they have there are overlaps between colvars and plume but the main important feature now that colvars has been more tightly integrated in grow max that you can that you don't have to hatch the grow max code to use colvars, but you use colvars from the native interface in grow max. I haven't tested or made any speed comparisons but I expect that it will make it more efficient than using separate library and with a patched version. Thank you. Thank you very much. So I will ask if you can kindly for colvars question just to refer to the forum, because I think it's more appropriate or not to the next webinar. Okay, thank you very much. So now we go further with the following question for Remy. I will allow you to talk. I think he has some, oh no, I cannot hold, sorry. With the deformation option from the MDP, how should we then extract the viscosity similar to NAMD simulation? So you can analyze it using the GMX energy tool and then there is, I think there are actually two options to calculate the shear viscosity and the recommended one is using the Einstein formula. And with that, you don't have to write the output very often but you need to at least do the energy calculations, frequently enough at least to capture these motions. Okay, thank you very much. So then we have Fatemej, oh, sorry, Fatemej Yul, I tried one more tool if you can speak, and then you can grant. Yeah, you can speak, we hear you. No, we just got the background noise. So the question is, is it possible to calculate the viscosity of aqueous system containing aggregates? Do you hear me now? Yeah, now it's we hear. Okay, yeah, this is my question. So because your example is for some homogeneous system, so what will be happen if we have aqueous solution containing aggregates with other molecules, is it possible to use the same approach to calculate the viscosity? I think if you have a more complex system, so in these cases, we also have mixtures, but so heterogeneous but still simple mixtures. If you, for example, would have more complex setups, if you would have, let's say, a protein in there or something, I think you would have more difficulties calculating the shear viscosity. Thank you. Okay, thank you. And we have a following up question is also on viscosity that I will read from Mattia. So, he would like to ask, well, where we could find more information on the details of the calculation of the viscosity using the cell deformation feature. Maybe it was mentioned during the presentation, but it didn't catch the information. Thank you. If I can't step in, sorry, there will be also a tutorial on this soon. Please go ahead. Yes, so I would refer firstly to the 2024 release notes where you can find a little bit more information on what has been changed. And then I would recommend you to read the manual, the Gromax manual where there is more detail about calculating shear viscosity. And also how these options work. So I hope that answered your question. Thank you. So, now we see, sometimes again, could you, I am mute you, could you state your question. So his question or her question is, is a non discriminate CV possible, such as the distance from an atom. From an atom B to an atom C, falling with a center range at a given time. I don't know. That's my answer. I would recommend you to ask that in the call bars session into weeks. It's not possible in Gromax without call bars or possibly plume. Okay, thank you. So now we have an order and I will try to unmute you if I found you so good. Please, not up. Go ahead. Hi. Yeah, so I actually, I was surprised that when I saw my name so thanks for using our system as an example. Yeah, thanks for the good reference data there. So my question would be that I don't know too much of this new feature of the age that you mentioned but this feature of scaling the target distribution. Would it be a good option as a default or is there some examples when it would be useful or could fail or so just to get an idea if if from now on that would be something to turn on. In general, it would be a good idea, but we want to be careful with turning it on as default. We have seen cases where you can get self reinforcement of the target distribution because often you see that when you sample a region more, you will get a slightly higher friction metric because it's based on the auto correlation function. And the longer sampling there could lead to higher friction metric, which means that you could you could in theory get an increased target distribution in regions where you sample more. So you might want to be a little bit careful with that and I should also add that you shouldn't expect any extra or any magic from this the scaling by default, or the scaling by this option only takes place after leaving the initial stage, where you have had enough sampling to get reasonably reliable friction coefficient. What you can do instead is also that if you have a friction coefficient that you trust you can also use that as a user input to assimilation. But in general, I would recommend experimenting with this at least especially if you have a large differences in your estimated friction metric from a WH. Okay, thank you. Yeah, that already is helpful. Thank you. Thank you. So I will go to Abishek that he has a performance question. This is not any more online. I cannot see any more online. Okay, I asked the question how much adding a simple center of muscle strain with coal bar will affect the speed. Here I don't know exactly either but I wouldn't say very little. That's my impression. So using a WH with a native grow max pooling compared to using a coal bars with a similar setup I didn't see any difference in performance on my local system here. Okay, so we have also another question from Khan will update on the GPU using verbert integration and the baby option also become available in the foreseeable future. I guess foreseeable future like within one year. Yes, there are currently no concrete plans for the and the velocity of early integrator on the GPU so I can't say anything for sure there. Yes, and he has also thank you. There's also another question have been struggling with the protonation of NH three plus without specifically choosing which hydrogen to force away from the nitrogen using blown. I guess this is also question for the form. Yeah, I think either the forum or the webinar in two weeks. Yeah, so then we have. I will take the last question from. I will see if it's still online. Yes, I try to unmute him. We'll see if he can speak. Oh no sorry told that the microphone is not working. So how much precise the H atom way transfer in, in case of membrane structure. I guess that is my guess. Yeah, I'm still not quite sure that I understand what is being asked here but in general, transferring mass from the heavy atom to hydrogen's will. Will not have a large effect on on the sampling unless you are really studying details about well rotational degrees of freedom, I would say there might be other situations as well but. But in general it's safe to transfer the hydrogen mass or mass to the hydrogens. Yeah. So, thank you very much. Thank you very much for all the question. Thank you very much for manuals for presenting grommets. And now I would like to feel just give me the control. Yes, I will just present the next webinar. So I just put it already in the chat. But so the next webinar will be in two weeks. I put in the chat the link so that you can enroll. And it will be on an unsampling collective variable use space using the cold bars library in grommets. And then we will have three speaker. Jacomo Fiorin, Humbert Sanzos at Jerome and in, and it will be the 19th of March 2024 at the same time. 15 and here I see that is, I think this will be not yet summertime it will be still winter time, but yeah, so but at three o'clock. Okay. Thank you very much for being with us and see you in two weeks. Bye bye.