Episode 34 - Rafael Bernardi: Molecular foundations of mechanostability, development of VMD and NAMD, and QM/MM interfaces

Transcript

Milosz 0:12

​​Hello. Here is episode 34 of the Phase Space Invaders podcast, where we explore the landscape of computational biophysics, and my guest today is Rafael Bernardi. Rafael is a professor at the Department of Physics at Auburn University, and his group is part of the famous NIH Center for Macromolecular Modeling and Bioinformatics. That also includes two former guests of the podcast, Alexei Aksimentiev and Zan Luthey-Schulten. Within the group, Rafael carries a large part of the responsibility for directing the development of the tools we all love, the molecular visualizer VMD, and the molecular dynamics engine NAMD. But aside from that, of course, he runs his own independent research agenda that revolves around questions of mechanostability of biomolecules, force transmission and resistance, allostery, and adhesion. So we start from discussing the applied research angle, talking about the molecular mechanisms of catch bonds, all the theoretical peculiarities of studying force-induced phenomena on the atomistic scale, and, uh, modeling results from atomic force microscopy, AFM for short. That serves as, as a segue into the question of software development and how important it is for our tools to answer a need that actually exists out there, especially when we do that at scale, as happens with both VMD and NAMD. We comment on the ongoing release of VMD 2.0, which I encourage everyone to check out, and, uh, the shift in philosophy that motivated this development. We end up defining, half-jokingly of course, a mini manifesto of contact-based biochemistry, a perspective on biomolecular events we both share from the points of view of our respective research programs. Plenty of cool stuff, so as always, enjoy our conversation. Okay. So Rafael Bernardi, welcome to the podcast

Rafael Bernardi 2:22

Thank you. Thank you for having me,

Milosz 2:25

So Raphael, your research in the last decade or so has been deeply embedded in the NAMD/VMD community, including major contributions and supervision, uh, to the software ecosystem. But I have to say I'm also quite passionate about your other research lines, such as, you know, nanomechanics, allostery, mechanostability, force transmission. I even have a hint that some of those fundamental scientific questions might go back to deep history when your postdoc supervisor supervised the postdoc supervisor of my PhD supervisor as they both, you know, Klaus, Klaus Schulten and, and Helmut Grubmüller laid kind of the foundation for the study of molecular mechanics in the most literal sense of the phrase. Do you mind sharing what brought you into this specific field? And, you know, it always feels like destroying molecules should be straightforward, right? But I assume we have actually learned something about that since.

Rafael Bernardi 3:21

Yeah. So, okay, I, I, I got into this kind of by accident. I came to work in Klaus' Lab in the US. So I'm originally from Brazil. I came from Brazil to work in Klaus' Lab study biofuels and second generation biofuels. So we want-- we were interested in studying enzymes that degrade biomass, split biomass. And in one of product, the most efficient enzyme that does that is a large complex called cellulosomes. And the cellulosomes have some interesting mechanical properties. So coming to work on biofuels- on enzymes that degrade biomass, and then cellulosomes actually got me in contact with Claus's friend, Hermann Gaub. And Hermann,, was interested in the cellulosome as well because they form this large complex that should be mechanostable. So Hermann, had a, a group leader at the time, Michael Nash, you know, like who was, interested in, in pulling on these complexes, and we start pulling on them together, kind of like trying to understand how we engineer these large complexes. And that actually got me into this mechanobiology because we started to study like, oh, this is quite interesting, like the mechanics of these systems. we end up collaborating like, Klaus, Hermann, and also Ed Bayer, who was the one who discovered the cellulosome. So like we had all these big names, you know, like interested in studying that. It was very nice. Like it was a very nice group, and we actually started to learn mechanics that people could not learn before, and that got me kind of hooked into this world of mechanics

Milosz 4:55

Right. So I assume this is something that's just a structural, element, right? That kind of conveys rigidity. Or wh- why is it, um,

Rafael Bernardi 5:03

Yeah. So,

Milosz 5:03

s- specifically

Rafael Bernardi 5:04

that we studied was the cellulosome. And, and I, I think what caught my attention, I think that what, like, made me so passionate about it is because it was definitely not obvious.

Milosz 5:13

Mm-hmm.

Rafael Bernardi 5:13

know, like, so the, the first part is this is a catch bond system, and a catch bond is a bond that,... Like, explain this shortly. Like, regularly chemical bond is what we call a slip bond. So that means that if you apply a higher force, they live less. So, like, if you apply a force, a little force, they might live five minutes. If you apply a higher force, they'll probably live one minute. Now, catch bonds are the opposite. If you apply a force to them, they live longer, so they, they kind of get hook. And, the cellulosomes end up being, uh, this, like, cohesion dock grains in cellulosomes end up being, uh, catch bonds. And we, when we were simulating them, it was the first time that somebody could show really, like, clearly atomistically the mechanism of a catch bond. a catch bond. Now, because, like, we have seen since then that catch bonds can be actually very different than that. But the first one was very interesting because we were pulling this system that has, like, a very flat surfaces. Like, basically it's a flat surface from the cohesion, a flat surface from the dock grain, and you pull them apart, the contact area goes up,

Milosz 6:17

Oh yeah.

Rafael Bernardi 6:17

Which is completely not intuitive. You know, like, you are actually pulling something apart and imagine that they are actually tight, getting tighter. And this is pretty much, like one of those Chinese finger traps. You know, like, that's typically what I use to, to, to demonstrate this the finger trap, basically, like, if you go very slowly, you can remove your finger very easily. But if you pull fast, you lock and it, it basically traps your finger, right? The finger trap. these, these molecules are doing similar things. They're getting trapped in those conformations,... And that is, like, so, uh, unexpected to me because I think, like, can think about large things and, and engineer a system that is like that, you know? Like, but, like, doing something at the molecular level that is doing the same thing was like, whoa, you know? Like, that's like, had this whoa kind of thing on me, and that's how I end up getting another one and, like, studying and another system. And now basically what we do is study those types of systems.

Milosz 7:10

Yeah, that kind of defies our intuitions, right? Because we maybe as simulation people or, mechanical statisticians are used to thinking of equilibrium properties that are path independent, right? But you have a clear path dependence here, and like you have to go to all those kind of non-equilibrium methods

Rafael Bernardi 7:29

simulations, yeah, it's, it's kind of like we are typically doing like SMD and those like steer pullings like that for calculating free energy. So we wanna go like at the quasi-equilibrium regime. Like in

Milosz 7:40

Yeah.

Rafael Bernardi 7:41

opposite. Like I wanna study the mechanical property. I wanna see what is happening when you give like a big, a big pull on those systems. And, and, and it's, it's very different. And, and it's, it's actually extremely biologically relevant. It's just that we don't study it so much because the tools were not there, both

Milosz 8:00

Yeah

Rafael Bernardi 8:01

experimentally, you know?

Milosz 8:03

Well, you would think that a lot of nature has mechanical properties. Like nature is pretty mechanical, but I would say there's probably very few specialized molecular components that actually carry most of the loads or do most of the adhesion or force resistance, right? So like in the whole biological, landscape or universe,

Rafael Bernardi 8:25

There's, there's much more than we even think. There's a,

Milosz 8:25

it's a pretty specialized corner.

Rafael Bernardi 8:28

there's much more than we usually think, you know. Like

Milosz 8:30

Oh, really?

Rafael Bernardi 8:31

Yeah. So there are-- Like we have been seeing more and more over the years. and some of the reasons are some of the system are super resilient, meaning like the, the tools that we developed-- Okay. Like you have to remember that pulling proteins experimentally was developed in the 1990s. So like these are like the first proteins being pulled.

Milosz 8:53

We're talking AFM setups, right?

Rafael Bernardi 8:56

like like by Herman Gaub, like his lab, you know. Like, so like this

Milosz 8:58

Mm-hmm.

Rafael Bernardi 8:58

this is kind of like some of the first experiments. And people doing also with optical tweezers, trapping molecules, late 90s. But like pulling, doing like single molecule force spectroscopy is, is relatively recent. To do that really well, to do that like kind of a standard technique is something from the last 15, 10 years. Now, going into high level forces, it's actually our-- One of our papers was the first one in 2018. know, like to go like beyond like strep avidin-biotin. You know, like going kind of beyond the, the normal strong. the whole-- Like you have to engineer the whole, uh, surface chemistry to be able to do it. And, and, uh, like all the experiments actually get more and more complicated. Also, there's the other side of it. A lot of these forces happen inside cells, and you do-- You wanna study them inside the cell. So there's a lot of people developing tools to study, at this like cellular scale, like very small forces. And so there's a lot of molecules that we just, we just don't know yet. there you see them popping, like in new conferences, in new, editions for like large journals, you know, like on mechanobiology. A lot of the molecular mechanobiology. There's a lot of coming up, uh, these days

Milosz 10:08

I see. Is it more like inside the cell or between cells in the sense of adhesion versus transport or, or both?

Rafael Bernardi 10:16

M-my group studies more the highly mechanostable, so they're usually outside the cell.

Milosz 10:21

Yeah

Rafael Bernardi 10:22

Uh, but a lot of the mechanobiology being studied at the moment is inside the cell. So these are like, like a few piconewtons of force, sub-piconewtons of force that make the whatever action you need

Milosz 10:34

Well, we just studied the ATP synthase, which is like a-a-another project that was, you know, shared interest between Helmut and Klaus, uh, if I remember correctly.

Rafael Bernardi 10:42

Exactly. Yeah, yeah

Milosz 10:43

I'm kind of familiar. But then, of course, it's like a molecular machine that has a timescale well beyond our simulations, right? So, like, the question is always to what extent we are capturing the relevant timescales, with the latest simulations. Is it already there to, to kind of see the biological scale or experimental scale, which can be also different, right?

Rafael Bernardi 11:07

So if you look at the ATP synthase, you are looking at like this rotational motion that happens in a much longer timescale than we can usually simulate. You can use, of course, some enhanced sampling methods and all of that. the point that I was trying to make is like an ATP synthase is gonna produce maybe a couple of piconewtons of force stroke to make that rotation. Like while if you go outside of the cell and you're looking about adhesion mechanisms, you are on the hundreds of piconewtons to maybe a couple of nanonewtons. So like it's like three orders of magnitude, sometimes stronger. Remember that to be that strong, you also, means that the, the lifetime of that bond is to infinity compared to our time of, on simulations. However, if you wanna study their properties when you're breaking them, simulations are pretty doable because you wanna keep pulling until they rupture. So you can actually make that much faster. Of course, experiments are not that fast, but there are theories that can connect the, the two speeds and you can study

Milosz 12:10

I remember a paper that essentially looked at streptavidin biotin and I think connected the two time scales, right? Uh, that was a couple of years ago that

Rafael Bernardi 12:17

a paper by Helmut Grubmueller, like the co-simulation part, and the experiments were done by Felix Rico. Yeah, they published a few years ago, I think it was a PNAS, where they connect the two worlds. actually have something on the pipeline very similar to that, where we are in many, many different-- Like we have 10 orders of magnitude different speeds on, AFM and, and, simulations

Milosz 12:43

Yeah, because that's also becoming another validation of, whatever force fields in our systems, right? And there are so many variables that you can get right or wrong. I know that you also looked at directional pulling, right? So like you can have different directions in which things are being stretched

Rafael Bernardi 12:59

Yeah. So one of the-- I think, okay, one of the consequences of the things that we have been studying is, okay, we started that because we wanted to understand discrepancies in the data. And the discrepancies here were if you look different labs, they were getting very different results for the same pulling. And sometimes even your own data from the same lab was getting very different results, and that was for streptavidin-biotin. And one of the things that we noticed is that, okay, that's kind of like discussed in literature, but not well. Kind of like people kind of know but could not prove that,, this was related to tethering. Like meaning like you're using, reactive like amines to do it. Like anything that is kind of-- any amine that is exposed, and then you hook there your streptavidin. The biotin is pretty well controlled where you're pulling, but the biotin binds to the streptavidin, and you pull. But now your pulling geometry differs depending on where you were hooked. So what we were able to do is both experimentally and computationally pull from different places, kind of create a way of knowing where you're pulling in the experiments. And then, of course, the simulation is very easy. You just pull wherever you want and then show that, different results were actually coming from pulling from different places. When you, of course, have like dozens of different places, this becomes a huge distribution. You don't notice that it's actually, many different places, but, you can actually get these different geometries causing a different result. Again, there is some counterintuitive results out of this, which is very interesting in my opinion. is streptavidin, basically locks biotin kind of like your hand going around another finger. And one of the geometries is the geometry that kind of forces this out and opens your hand. Kinda force the biotin opening your hand, opening the streptavidin, right? And you would like-- at least myself, I would initially assume that, okay, this is the strongest one because the other one would be kind of like taking your fingers straight out, you know, like without destroying your hand, without destroying the streptavidin. But actually, peeling is cheaper than breaking many bonds at once.

Milosz 15:10

Yeah

Rafael Bernardi 15:11

that you go to the side and peel the streptavidin off, like basically break the whole streptavidin, is weaker than the one that if you pull just straight out. think those kind of-- when you think about it, yeah, it makes sense because you're breaking like 10 bonds at once instead of breaking one at a time. it's, uh

Milosz 15:29

Right. Here's where the path dependence comes in.

Rafael Bernardi 15:32

Yeah.

Milosz 15:33

And, uh, the way-- Of course, you're, you're like you're showing things with your fingers, which is not going to be captured by a podcast. But,

Rafael Bernardi 15:39

I know.

Milosz 15:40

but I remembered, uh, immediately this idea of the haptic device from VMD, right? Uh, so is, is something like that actually being useful when looking at, responses and getting intuitions about forces? I mean, I, I'm referring to this because, you know, you're also a big part of development of VMD and many, uh, tools that are being added on top of that.

Rafael Bernardi 16:05

No, this is, you know, a lot of the times these kind of tools like the haptic and, and a lot of these gimmicks are pretty fun, but I would say they are not so useful on a daily basis, you know, like uh, it's, it's not a major effort on development, I would say. We're not doing anything on that direction at the moment.

Milosz 16:24

Right. But how much of a help is it that, you know, you're doing your essentially research agenda together with software development, right? Because you can kind of steer many directions in which both VMD and NAMD are going and, uh, is that helping?

Rafael Bernardi 16:43

it, it does, of course, because, when I was in, Klaus lab, because of the funding mechanism we had at the time, way the biology was treated was what we call this driving biomedical problems. So we had driving biomedical problems that would, push the boundaries of what the, the computational needed, and then we would develop softwares to kind of accomplish that. the idea basically remains in my lab. You know, like I, I like that idea. I like the

Milosz 17:14

Yeah

Rafael Bernardi 17:15

we are pushing the software development because we need it. So in regard, there are a lot of tools that come out in VMD sometimes because we need them. You know, like there is like a... The VMD 2 has a huge suite of understanding, contacts and hydrogen bonds and how you analyze them, how you make con- like networks of contacts. All of that is something that we have been doing in our lab for a long time. We just put it into a nice tool inside VMD to speed up things in our lab, uh, especially training new people, you know. Because the, the, old people like myself are not learning any of those tools anymore. You're basically writing scripts to do it. the, the younger people that get into the lab, it's, it's a much faster path to learn how to do it. It honestly is a much easier tool. You know, like, I think it's, uh... What we do then is to create these tools that help people to do the analysis that we are doing. in- inside VMD, I would say these contacts and networks and all of that is getting more and more space because we use them a lot. way for free energy methods and all of that. They are like becoming more part of uh, at least on the VMD side. On the NAMD side, I can tell you that there are some pushes too, like, SMD was one of the first tools to become multi-GPU resident, you know, like in NAMD. Because we used it so much that we wanted that. there are other developments in NAMD that are kind of pushing the same direction. Like, we recently needed more, ways of keeping very, very long simulations. We had to make a few changes in- inside NAMD so that we would not break anything because we're going like through many, many, many microseconds of simulations pulling, and so that was breaking some stuff.

Milosz 18:58

Hmm. Yeah.

Rafael Bernardi 18:59

that's all. when you create this tool, sometimes you don't imagine that people are gonna pull for 10 trillion

Milosz 19:06

Yeah, yeah, yeah. This is the common thing that in Amber you, for example, you set the number of steps and it's stored in like a standard integer, which can only be like one microsecond

Rafael Bernardi 19:17

Exactly.

Milosz 19:17

and it always breaks something on a technical level. Yes

Rafael Bernardi 19:20

Yeah, but usually you solve those problems with workarounds very easily. But when you're pulling on the system the way it's set up, you actually cannot, you know? So we had to fix some of that.

Milosz 19:29

I think it's a common, common feeling that, a lot of time gets wasted on solving the most trivial questions, like formatting of PDB files, right? Or something doesn't go in the field that only stores four decimal points or whatever it is. Yeah

Rafael Bernardi 19:43

that is part of the, discussion that we we have all the time in the center, like in this development of NAMD and VMD, that a lot of the, the time we spend is actually on developing things that are, very hard to, to justify and write on a grant, know? Because-- But You

Milosz 20:01

Yes. Yes

Rafael Bernardi 20:02

those are needed, you know? Like it's just that is, it's not very sexy when you're writing on a grant, you know? But it's very hard to-- But you need them. It's a, it's important thing

Milosz 20:10

That's true. Yeah, I had this, discussion recently with someone who was complaining about like, you know, you cannot write a grant for a method essentially, right? So I was explaining this perspective that you were just sharing that, okay, you have to have a bigger goal, which is usually tied to something people care about, and then you can sweep everything technical under the rug essentially, right? But that's, that's kind of the way we've been doing this, um, that,

Rafael Bernardi 20:32

Yeah,

Milosz 20:32

you

Rafael Bernardi 20:32

it's

Milosz 20:33

we need the things and we, we, we find excuses to do them.

Rafael Bernardi 20:36

It, it's, an unfortunate thing, you know, like the, the funding mechanism, of course, is driven by the science we do, which, sure, for one side makes sense, but the other side is very hard to justify the tools. Even though we are developing tools for a VMD, we have with a very modest type of counting 380,000 users, know, like earlier. So it still is hard to justify, you know? And

Milosz 21:02

Yeah, this is pretty striking always that, uh, last, uh, episode we had a conversation with Wonpil Im who's running CHARMM-GUI, and he mentioned that, you know, CHARMM-GUI is just like a single server with two GPUs despite the number of users that they have. I'm like, there's really a big mismatch in how much money we spend versus how many users something might have. VMD is another clear example, right? That

Rafael Bernardi 21:24

Yeah,

Milosz 21:25

it's the same-- In the end, it's, it's probably mostly three people working almost full-time, but not, not exactly full-time on, on making this happen.

Rafael Bernardi 21:33

I mean, we are lucky to have this big NIH resource funding that supports NAMD and VMD and some of these other tools that we develop. but, the reason why we have them is, in my opinion, is because we always try to show economies of scale. basically, like develop these tools for hundreds of thousands of people, and it costs NIH much less than having each lab developing their own small tool, it's much

Milosz 22:00

Right

Rafael Bernardi 22:01

The part that bothers me about that particularly is, take NAMD. NAMD is, in many supercomputers in the US, the most used software, like the software that spends the most hours on, running jobs. especially in com- supercomputers that are more targeted for large systems where people are doing like whole cells, like whole viruses and things like that, at that scale. So let's take that a com- any supercomputer like that costs anything between million and, and like a billion dollars, like in investment from like DOE, from NSF If we can make a software run 10% faster, that computer could be 10% smaller or, or maybe, you know, like could do 10% more with it.

Milosz 22:46

Yeah

Rafael Bernardi 22:47

I think a lot of the times, the funding mechanisms don't think about like, okay, if 30% of the time the supercomputer, 20% of the time supercomputer's been running this or that software, of that money could go to that software development to make the software run as fast as they can on that, device. Like I said, uh

Milosz 23:05

But then there's this, you know, there's this paradox. I forget whose paradox that is, but essentially says that if you make it faster, people will use it more because now you can do 10 replicas rather than one replica. So in the end, yeah

Rafael Bernardi 23:19

the unfairness of the competition. but in terms of, money well spent, I would say it's

Milosz 23:24

Yes.

Rafael Bernardi 23:25

reach.

Milosz 23:26

No, no, definitely. I'm just joking here. But, uh,

Rafael Bernardi 23:29

Sure

Milosz 23:29

also realized that the more we have-- the faster the simulation, engines and so on, the more we actually simulate, right? It's not that we simulate the same amount we just find more reasons to, to run it 10 times and see which, run is, you know, more interesting or whatever.

Rafael Bernardi 23:45

I, usually joke in my lab that my thesis that got even an award can be done now in a day.

Milosz 23:52

Yeah. Yeah, that's, that's true for most of us now. Yes.

Rafael Bernardi 23:55

Mm-hmm.

Milosz 23:57

but then, yeah, so let's first get the word out because I, I'm not sure... I know you did a lot of effort to make it, sound in the community, but, VMD 2.0 is out now, and people can, you know, install it and test it. I've been testing it for a while, and it's really a major improvement in terms of, comfort and, uh, usability. but there are some major, developments. I know y-you are the one developing QwikMD, for example. I was just mentioning to you before we started this recording that I was very surprised that it's, like, one of the few pieces of software that I just opened up. I had a QMMM calculation to do, so I opened up QwikMD, and I set up the thing in a kind of quick and dirty way, and it actually worked, which almost never happens to me. So that's an impressive, feat. And I have this feeling that, you know, VMD is full of plugins and add-ons that few people know about. Uh, so I'm thinking, like, do you have a pathway to, change that?

Rafael Bernardi 24:55

I have the same feeling, and that's, one of the main reasons why... Okay, since I joined Klaus Lab, I was always vocal about changing VMD. You know, like I, I was-- I, I had this conversation with John Stone a million times, you know, like, we need to improve our interface. and I think there was a difference in philosophy, to be honest. I think there was a big change in philosophy for both NAMD and VMD. the NAMD philosophy was always we are serving the top 1%. You know, like we're serving the people that are running the 100 million atom system. And the VMD philosophy was ra- always like we, we run fast and we can do crazy analysis for the large systems, and we don't care so much about most of the... other users are basically like, uh, an add-on to this main development. And my philosophy is more like, we need to embrace all the users. And, and we need to be able to embrace users that are like not so keen on TCL. I think there's less and less people keen on TCL

Milosz 25:52

That's a big point there, yes

Rafael Bernardi 25:54

So we made this whole menu idea. So the whole menu of VMD2 was designed in CorelDRAW, like designing like buttons and like how do we click on it. And by the way, QwikMD was developed exactly the same way many years before. and the idea is like can we-- how do we make these buttons easy? How do we make this easy for people? And we changed many times over the last year, you know, like kind of like how we make this easier and easier and easier. And of course, there was a lot of things being developed under the hood. So, kind of to your point back, we were trying to make the things that we find more useful for most people now very visible, like Covars module, you know, like QwikMD. things are now in a clear button right away. And we are trying to make them all the analysis tools in the same direction. So like instead of having like, okay, the one that always annoyed me, three RMSD tools.

Milosz 26:42

Oh, yes

Rafael Bernardi 26:43

have one tool that does more than the three together before, you know. Like, so we try to make this, instead of a, a hydrogen bond, we have now one that calculates contacts, hydrogen bonds, all of that in one big tool, kind of try to group things that are somewhat similar. And so the users are gonna be able to find them more easily, and they're gonna discover things more easily. I think that's, that's some of the ideas behind it. but there's always gonna be hidden things. Not hidden because we wanna hide them. It's just because, you know, like we cannot make a menu for everything. So like you're gonna have to type a command to launch them. That's, that's the

Milosz 27:16

Yeah. I'm, you know, I'm guilty 'cause there was MovieMaker and now there are two MovieMakers in VMD, so I know how these things come to life. It's like a few people make their contributions and then nobody's really in charge of merging them together and maintaining them a long time, right? So that's the headache of the, eventual developer who takes charge.

Rafael Bernardi 27:35

Yeah. It's, it's sad because you wanna contribute. I think most of the user... A few, quite a few for these users that want to contribute and send something to us because a lot of it-- First of all, a lot of the times, whatever is being created already exists. because it's hidden somewhere. Um, but also i-it's very hard for us to keep everything in the same standard, to keep everything working the same way, you know, to keep things looking the same way. That's something that we are being very careful now with VMD2, trying to have a standard. But Diego is working very hard on creating a standard, so people when they want to develop, actually follow the standard. That's another thing that we are coming up with, you know, like for the next few months.

Milosz 28:18

That's gonna be very welcome. 'Cause yeah, there can be a core ecosystem and a kind of additional, right,

Rafael Bernardi 28:25

Yeah.

Milosz 28:25

uh,

Rafael Bernardi 28:26

we are

Milosz 28:26

ecosystem of tools

Rafael Bernardi 28:28

There's a lot of changes also in our paradigm. You know, like, like we used to use a, a CVS, like controlling version system, like inside the lab. Now we all use it, like for NAMD, we are using GitLab for some time. VMD, we have migrated already like three or four times to GitLab, but never actually make the full migration. hopefully do-- we are gonna be done with it this month, so that, uh, people can also contribute through there. And we are moving all our manuals, all our tutorials, everything to Git so that, uh, it becomes more water. I know this not, does not sound very impressive for most people because, you know, like people have been using these tools for a while. But one of the things that really impressed me about, uh, NAMD and VMD, like about Klaus Schulten lab, like when I joined, is actually a lot of the tools that we are using today, like things like Git, were in place there, kind of like a version of it, something like that was already being used there.

Milosz 29:20

Yeah

Rafael Bernardi 29:21

there was something that looked like a Dropbox that they developed in the 1990s inside

Milosz 29:26

Wow. Wow, that's impressive

Rafael Bernardi 29:29

VMD could do things like a Dropbox. I could send you a file through VMD. know, and nobody knows that. You know, those things were there. You know, like it's because they developed it when these other tools did not exist.

Milosz 29:42

That's probably where the joke comes from that, you know, eventually every tool you make starts sending emails, right? Suddenly becomes... I have the same problem with my libraries that eventually they become just a kind of black hole for every idea that they ever have. So

Rafael Bernardi 29:59

Yeah, yeah. So

Milosz 29:59

someone has to be in charge

Rafael Bernardi 30:01

And it's very important to clean things up, you know? Like

Milosz 30:04

Yes

Rafael Bernardi 30:04

I, uh, in, in a note for people thinking that, because we are the NAMD community, we don't-- We fight with, let's say, the Gromacs community. You know, like one of our main contacts and advisors sometimes, like people that we chat about stuff, you know, like I think we exchange, things, it is, Eric Lindahl from the Gromacs community. And one of the things Eric was telling me was like very strong about that. there is nobody responsible for the, that part of the code, part of the code should get out. You know, like it's, uh

Milosz 30:36

Yeah, it's not very easy to get into the GROMACS code base 'cause they have very strict requirements for that. That's true

Rafael Bernardi 30:42

And that's something that they had to develop over the years. It was

Milosz 30:46

Yes

Rafael Bernardi 30:46

20 years ago. And

Milosz 30:49

all the tools that were broken kind of a decade ago, maybe just because of that, just because someone contributed it and it rotted away over time

Rafael Bernardi 30:58

Yeah, it's a-- And a lot of the times that's why the development of these top codes is so much slower than the community wants because, you know, like for us to keep our sanity,

Milosz 31:10

Yes

Rafael Bernardi 31:11

is, a lot of work and a lot-- it's, it's quite slow

Milosz 31:15

And then there are so many dimensions, right? Because you are also responsible for implementing a large chunk of the QMMM interface in NAMD,

Rafael Bernardi 31:25

Yeah

Milosz 31:26

supports I don't know how many codes now. Uh, I know that there are two that are kind of main. I think it's ORCA and MOPAC that are the main connections, right?

Rafael Bernardi 31:35

we, we developed it around ORCA and MOPAC. So the, the whole story is that we had, we established collaboration with groups doing QM. Like, you know, most of us, like me, pretend we know some QM, but you know, like we are not QM experts. So we collaborated, with, the people that like from the ORCA community and the people from the MOPAC community to kind of, create a tool. But we also wanna make sure that it was available for any software, and they were very helpful in guiding us how to make that. So the MOPAC idea was like, let's make a connection to a semi-empirical that is as fast as possible so that a lot of people can do it. And ORCA was like, okay, I wanna do something with more precision, with like higher level accuracy. ORCA has a lot of these tools, that's how we, we made NAMD. but we also created this, descriptive interface, basically. Like it's an interface that sends information in a specific format, gets information in a specific format back, and you can write a very simple Python script that can connect to absolutely any software. And we, provided some exemplary scripts for, and, and, and Firefly, I think GAMESS, uh, TeraCam. You know, we provided for some of them and people could build on that. And we got people building all kinds of stuff, you know, like including like ANI, you know, like doing AI for base to QM, and that's all possible through NAMD, you know.

Milosz 32:54

Yeah, I have to say that it was really very easy to set up. So I recommend to everyone who's got their like tiny QMMM project, shelved for later, you can go ahead and try to set it up. It's, it's pretty straightforward, gotta say

Rafael Bernardi 33:09

And, also another thing is, that-that's my vision as a biophysicist, not as a chemist, you know. Like not a chemist. I cannot pretend to be one. I, I think as a biophysicist, I tend to think that, QMM is very important in our field. But in biophysics, you have to remember that, entropy plays a large role. So like you need to have dynamics, you need to simulate for a long time. And that is why I think we developed NAMD around that philosophy, that you could also connect any of these tools to our enhanced sampling methods, to the free energy methods. So you can do a string calculation, you can do metadynamics, you can do,, ABF. You can use all of those things together with QMM. So you can actually get a very nice free energy profile of a, a chemical reaction. I think that is a huge value compared to, uh, many of the other, uh, QM softwares and like QMM softwares

Milosz 34:05

Yes. Again, that's what I did. I did, like a metadynamics of a proton transfer in MOPAC, and it was the first time I used Colvars. It was the first time I used, NAMD actually in a long time and it all kind of came together. So

Rafael Bernardi 34:17

Yeah, yeah

Milosz 34:18

I attest it works. Uh, did you have to by chance ever use QMMM for like mechanobiology or is it just enzymatic, uh,

Rafael Bernardi 34:28

have,

Milosz 34:29

reactions?

Rafael Bernardi 34:30

we have, we have done it. Um, we have a system where polarization plays a huge role. So, we have this, one of the main things we study the, the lab is adhesion, like in, in bacteria, and those are, like, very important for staph infection. And, the adhesins are actually the most stable mechanical proteins, uh, ever found, the mechanical bond ever found. they also have, like, so the adhesion part is what you call the A domain, but they also have this, like, chain of B, what we call B domains. And these B domains are basically like shock absorbers. They break before you break the main contact, so you don't like, you never lose your grip. And they are the most stable protein folds we know under force. And they do that by having three calcium on forming a crazy loop around just a regular Ig fold.

Milosz 35:15

Oh well

Rafael Bernardi 35:16

fold can break in a few piconewtons. This one goes through nanonewton. And, so these calcium loops are like three calciums being coordinated by some amino acids. Like in that case is where Classical MD does not work so well. we, got to do with QMMM. The problem with QMMM is so slow that we had to go so fast a pull that things were not really working the way we wanted. We, curious enough, we end up finding that we got very nice results when we do, uh, polarizable force fields. And

Milosz 35:46

Yeah, that makes sense

Rafael Bernardi 35:47

yep, and then we got to do-- We did it with amoeba in OpenMM, and we did with, uh, Drude in, in NAMD. the results were very similar, actually. So we expected that amoeba would be better because of a better quadrupole, precision. But, we got very nice results with Drude, and then we end up, that end up motivating back to your question from a few questions ago, that end up motivating also the development of, the GPU resident version of Drude implementation in NAMD. So now you can

Milosz 36:19

cool

Rafael Bernardi 36:20

do Drude fast

Milosz 36:23

Oh, wow. That's useful. Yeah, I remember some-- probably was, uh, Frauke Graeter doing like bond breaking in extreme nano,

Rafael Bernardi 36:33

Oh,

Milosz 36:33

uh,

Rafael Bernardi 36:33

yeah.

Milosz 36:34

mechanosensation, but I think that's a very rare case, right? When-- So, but it's also interesting to see that you've, managed to, to connect those two worlds where you might need to use, well, in the end polarizable, right? But,

Rafael Bernardi 36:46

Yeah

Milosz 36:47

but still beyond, additive force fields for mechanosensation or mechano, sensitivity

Rafael Bernardi 36:55

Yeah, there, there is a lot that can be done with, sometimes simpler tools than we want. like the QMMM sound like the best thing to do, but then of course the limitations. And like I think that's something that I always try to tell people that have to use the tool that can do the job to answer the question that you want, because, uh, computational work is very expensive. need to remember that a lot of the times the work we are doing is more expensive than experiments because of

Milosz 37:21

True.

Rafael Bernardi 37:21

That's all the...

Milosz 37:22

Even though we claim otherwise

Rafael Bernardi 37:24

we just forget that the computer costs money You know, we always need to be careful about that, you know. Like it's, it's... Think about before you actually launch something, like think about what is the best way of achieving the result

Milosz 37:38

Yeah, that's a, I think a big question whether, you know, this whole machine learning, revolution will actually result in better semi-empirical force fields. There are some that have been released recently and have big claims. I haven't still tested them, but, uh, you know, people are thinking, okay, maybe you're gonna be able to do QMMM for a fraction of the cost with way better accuracy. But I think that remains to be seen or maybe you have some opinions.

Rafael Bernardi 38:03

I've seen this as like a-- There are different philosophies on how you do that, right? As far as I understand, like the, the QM part. because there, there are people that are basically guessing the results with the AI, and there are people that are guessing the results of some of the calculations with AI, you know. So you still have your QM, uh, matrices, your QM calculations the same way, but instead of s- calculating every integral, you guess the result of an integral. That-that's one approach. the other approach is basically guess the result already directly, based on the training, of course. the results that I've been seeing are pretty, pretty good, you know. Like, ANI for instance, all the results I see with ANI, like from, Brett Berg's lab at Florida, it's like always very, very good.

Milosz 38:42

Mm-hmm.

Rafael Bernardi 38:42

but I mean, I'm, I'm not really in the field to be reading much about it, but everything that I see is, is pretty impressive. I think has a big, role in the, in the future.

Milosz 38:51

More work for the QMMM interfaces

Rafael Bernardi 38:54

Sure, sure. Yeah. I think that they become faster, we might also have to rework a whole interface because the whole interface was built in the principle that the QM calculation is so slow that we don't care about the communication being slow. Um, but you know, like if the, the, the QM calculation is as fast as the MM calculation, then the whole paradigm needs to change. You know, like you need a different way of connecting and all that.

Milosz 39:18

It's like the Born-Oppenheimer approximation of QMMM, right? If the collecting degree of freedom is way slower, then you can decouple them. Yeah, I see that. That makes a lot of sense

Rafael Bernardi 39:30

I do think there is a lot of place for AI on driving the MD though.

Milosz 39:36

Mm.

Rafael Bernardi 39:37

I think, that is something that coming-- It's, it's already here, but it's coming our way more and more with time. Not--

Milosz 39:45

You mean defining collective variables or, accelerating the simulation actually?

Rafael Bernardi 39:49

would

Milosz 39:49

Which way?

Rafael Bernardi 39:50

the simulation. But the collective variable, sure, that there are some tools on that direction already, um, that seem very nice results. I do believe that if we have enough training data, we would be able to do, do like microsecond long jumps basically, you know, like, like on how this is gonna be in a microsecond and force the system to go in that direction. And with that, we might be able to study processes that are from the micro to the m-millisecond more easily

Milosz 40:15

Yeah, like an implicit Markov state model almost, where you can...

Rafael Bernardi 40:20

but not necessarily calculating the minimum states, but actually kinetics, you know, kind of going through

Milosz 40:27

Oh, yeah.

Rafael Bernardi 40:29

Yeah,

Milosz 40:30

Well, they can always extract it from existing trajectories, but I think the hardest part is doing this through extrapolation from what they've seen, right? So, uh, but maybe coupling with some sort of structure prediction tools might work

Rafael Bernardi 40:45

Yeah. So in mechanical biology last year, what-- we published a paper where, where we were pulling on a system we then we were using our, our network analysis tool to calculate, basically can we predict the force based on the network of the rupture force? And we trained, of course, in one system, we did like hundreds of simulations, then we also tried to see how much in the past can you go and still predict. Basically, like how much in the beginning of the simulation it can be in this pulling simulation and predict what the rupture force is. And the results were pretty impressive, that we could predict the force well from like pretty much the beginning of the pulling.

Milosz 41:25

Cool. Cool. So there's like early something committing to the

Rafael Bernardi 41:32

about like this is

Milosz 41:32

outcome early on

Rafael Bernardi 41:33

being committed to become like a catch bond because it was a catch bond right, right on, you know? Like it was not something that happened later. So that,

Milosz 41:41

Yeah

Rafael Bernardi 41:42

pathway becomes committed very early

Milosz 41:44

I should apply it to some, uh, research ideas I still have in my head but something for another conversation. well, but overall, you know, my impression has been that you subscribe to, to what I call the, the contact-based, view of biochemistry, right? So this idea that most of-- 'cause, like, there's always this idea that things have momentum or, or things move through space. But in reality, all this time that molecules spend waiting for something to happen is just like for the right contact to happen and maybe a chain of contacts to form or, or dissolve, right? I, I don't know how much this kind of simple picture explains most of structural biology,

Rafael Bernardi 42:26

I think, I think my main vision is very similar to that. It, it just go one step beyond maybe. That is, the idea of using networks has become one of the major in my lab. Um, and the network here, dynamical network analysis, what that means is you've got to get the contacts, exactly what you're describing, but also the correlation. So in, in my view the following, it doesn't matter if they're, you're seeing is a hydrogen bond, is a salt bridge or a hydrophobic contact, whatever it is, or whatever contact that we don't-- we, we cannot-- we don't even understand well how to describe. If things stay close to each other and they move together, they are in contact, they are in a bond, that's kind of how I treat this system. So like I don't, I, I don't care what type of bond it is most of the times, I just care that this alanine is making a contact with this other, I don't know, serine, and they are always together and they move together and no matter what happens, they are together. So like what type of contact exactly sometimes, in my opinion, less relevant. And by having this vision, we have been able to engineer proteins like mechanical stable proteins much better than we could with just basic biochemistry principles. 'Cause basic biochemistry would never make a mutation like we did from an alanine and say like, "Oh, we're gonna make the protein much stronger." And we were actually able to show... there's a paper we published a few years ago, 2.6-fold stronger mechanical fold, uh, on a system proved experimentally based on a single mutation like that from an alanine to a glycine.

Milosz 43:59

Nice.

Rafael Bernardi 44:00

So you,

Milosz 44:01

yeah.

Rafael Bernardi 44:01

Don't predict those things based on just basic biochemistry. This contact idea like you're describing, the idea that they are moving together is actually, uh

Milosz 44:10

Yeah, it's like partitioning a molecule into a set of interacting rigid bodies essentially, yeah

Rafael Bernardi 44:17

Yeah. And, and if, and if they're interacting, like we are happy. If they are not interacting, this is like we can be improved. If you're an engineer, that should become stable or you can engineer the ones that are super stable, which become less stable, and that's kind of the, the idea

Milosz 44:30

Yeah, I guess maybe back to the connection with visualization. There's always this problem that when you look at a new system and you're like, you don't have a mental representation of what's important, what's not, and you run a simulation and there are like things happening, but it's so hard to wrap your head around what is actually happening. But if you partition it into meaningful bits and pieces, not necessarily a domain, because a domain doesn't necessarily have to be, you know, the single unit, but like sub-domains or whatever, it suddenly becomes-- I mean, this is something we're just observing now, doing a project here. But like it becomes so much clearer to observe what the changes actually mean and what is changing, with time or with, like addition of sub-components. So I think that's a really cool, kind of research project to, to push this vision forward as, you know, more visual and more kind of exploratory

Rafael Bernardi 45:23

Yeah. we have been trying to develop tools to help us inside VMD to do things like that. Like I say, it's very hard. Also, the other thing that is very hard is to bring in statistics, you know, like because we do many replicas,

Milosz 45:37

Mm-hmm.

Rafael Bernardi 45:38

do you create-- how do you visualize not only one network, but you did 100 replicas, so you're now gonna have 100 networks. How do you do that at once? Or how do you put weight on like, okay, this is, this is strength, but with this error bar, you know? Like, so it becomes super hard to do those things. We, we have been discussing how to do things like that for some time already, you know, especially because of the number of replicas we do in the lab. basically any paper that comes out of my lab has at least 100 replicas So it's a

Milosz 46:10

that makes sense for almost like, uh, Jarzynski style averaging in terms of,

Rafael Bernardi 46:17

we, yeah, because we are doing

Milosz 46:18

yeah

Rafael Bernardi 46:19

on the computer, you know, like so

Milosz 46:20

Yes

Rafael Bernardi 46:21

Uh, and that is something, the discussion that I had many times with Galb. And, and Galb was-- started to joke that he was retiring at the right time, you know, like, because the computers were catching up. Like the first papers on SMD from Schulten and from Grubmüller were, doing, like one, two, three replicates, know, like, and then you get this point. But now I frequently show in my talks that if I randomly pick like three points, you can get a completely different picture of what's happening in the system. Like you're comparing a mutant versus a wild type because there's a distribution of these ruptures, and this distribution is quite wide. And when you get like 40, 50, 100 replicates, you actually see this distribution the same way you see in the experiments. In the experiments, they do like 1,000 pulls, you know, like, and then you see this beautiful distribution. If we do 1,000 pulls, we see a beautiful distribution way

Milosz 47:10

That's a great teaching point that actually, you know, now we can also do that with simulations, right? Imagine what would someone conclude if they saw this simulation run for one nanosecond instead of five microseconds, right?

Rafael Bernardi 47:22

Yeah. So, a-and that is the thing, you know, like, uh, that's also a choice of how you do it. it better-- And like, really, I'm not giving you saying one is better than the other, but like, is it better to run 10 replicates at, let's say, 100 nanosecond long, or it's better to run 100 replicates at a ten nanosecond long. one is better? And the-- I think it really depends on the question you have. So when you're pulling on a system like us, when you're feeding that into, like, these, uh, dynamics for a spectrum kind of, curves that you see in experiments, what we have been seeing is that is many of the times it's better to go pull faster, but pull more times than pull slower and having fewer replicates. But again, it's the type of application we do. If you wanna study, like, some mechanism that takes longer to move, it might be the opposite answer. So it's always important to think about the question well before you start a simulation.

Milosz 48:18

Yeah, I think that was also the bottom line in like non-equilibrium alchemical free energy simulations, right? They do a lot of pullings, which are essentially the same concept of just pulling through a sort of collective variable, and they do it like hundreds of thousands of times

Rafael Bernardi 48:34

It's,

Milosz 48:35

get, uh, the distributions

Rafael Bernardi 48:37

Exactly. I'm afraid when I, when I see a lot of the times people coming and, and I talk to people that come sometimes for workshops, sometimes for, that talk to us on, on, on conference is that, people are expecting a, clear recipe. You know, like I'm gonna run-- If I run for a microsecond, I'm gonna pa- publish my paper in "Nature." You know, like that's not like that. You know, like sometimes one nanosecond is enough, you know? Like it's just like you, you need to be enough for the question you're asking, you know? Like it's all about the question. Like I have colleagues here that have attosecond lasers, you know? Like, so things that happen at the attosecond scale, you do not need a microsecond of simulation, my friend. Like it's, uh

Milosz 49:16

I was recently making this figure kind of, you know, showing what are different scales for different, uh, things that happen in simulation. I was always afraid that someone's going to take it too literally and, you know, will come complaining to me that, "Oh, I run my simulation for this much," and, you know, this never happens. But then, like, you take-- you have to have this flexible mindset that, yeah, things can take ten nanoseconds or ten microseconds, and you never have a guarantee until you, you see

Rafael Bernardi 49:43

The nature thing that I joke is because I've seen people like, like from my own lab, you know, like come in like, you know, like they published this paper in this very nice journal and, and, and they, there's only like 10 nanoseconds of simulation." I'm like, "Was it enough to answer the question?" know, like,

Milosz 49:57

Yeah,

Rafael Bernardi 49:57

it

Milosz 49:58

that's a

Rafael Bernardi 49:59

to answer the question. Well done.

Milosz 50:00

fair, question to ask. Exactly. If, if there are two ions in a small box of water, maybe that's all you need

Rafael Bernardi 50:08

And that's, uh, maybe that's all you need Exactly

Milosz 50:11

Okay, wonderful. Rafael Bernardi, thanks so much for the conversation, for the, you know, perspectives, a lot of insight into how things are developing and, uh, well, the about VMD 2.0 coming out soon, even stronger and better. I mean, it's already out there, but it's getting better every

Rafael Bernardi 50:28

Oh yeah,

Milosz 50:29

month, essentially

Rafael Bernardi 50:30

hope to have 2.1 by the end of the year. Let's put it that way.

Milosz 50:33

Okay. That's a big promise. We'll check in after a couple of months. Yeah. Thanks so much for the conversation. Hope you have a great day

Rafael Bernardi 50:42

Thank you all for listening. Thank you.

50:45

Thank you for listening. See you in the next episode of Phase Space Invaders