Phase Space Invaders (ψ)

Episode 2 - Max Bonomi: Interoperability, AI-generated ensembles, and recognizing all contributions to science

• Miłosz Wieczór • Season 1 • Episode 2

Share episode

Send us a text

In the second episode, Max Bonomi and I discuss efforts to achieve interoperability and portability in the computational community, and how the AI structural revolution will unfold to bring us realistic molecular ensembles. We then move on to ask how the entire range of contributions to science can be recognized, particularly at the early career stage.

Milosz:

Welcome to the Phase space invaders podcast, where we explore the future of computational biology and biophysics by interviewing researchers working on exciting transformative ideas. Today I'm talking to Max Bonomi a group leader at Institut Pasteur in Paris, who has been deeply involved with some of the foundational tools and methods for molecular simulations and integrative biophysics he was one of the key figures behind metadynamics and metainference, two widely used advanced simulation techniques that are reshaping our understanding of complex molecular systems. Max's work, particularly with the Plumed plugin, has also empowered researchers worldwide to modify molecular dynamic simulations with unprecedented ease and flexibility. On top of that, he is great at bringing the computational community together through workshops and conferences, and I have to credit him with inspiring me to start this very podcast after several wonderful meetings he co organized last year in the conversation, we exchanged thoughts on the future application of AI models to structural data and how we will have to integrate experiments, deep learning approaches, and physics based models to understand biology at a deeper level we also discussed the questions of credit attribution in the various scientific undertakings from peer review to event organization to code documentation trying to answer how we can ensure that every effort is valued in our scientific ecosystem. So without further ado, let's get started

Miłosz:

max Bonomi, welcome to the podcast. A pleasure to have you here.

Max Bonomi:

Thank you for inviting me. Thank you.

Miłosz:

So for context, I got into, into molecular simulations in, around early 2012. And, uh, already then it felt like somehow your name was popping all around the place. So I remember working with Plumed very early on and, uh, head around all the different flavors of Meta dynamics and then at some point came meta inference. And so the whole field kind of shifted towards, integrative ensemble modeling. And what I'm saying is, uh, from my perspective, these contributions felt like, uh, threshold moments for the community. Do you mind telling us what got you engaged in so many meta-level projects,

Max Bonomi:

Ah,

Miłosz:

say?

Max Bonomi:

good. Good question. so maybe the first thing I, I should mention is, is where all started, which is the development of Plumed. And this dates back to the time of my PhD in, uh, ETH, in the group of, uh, Michele Parrinello, who was very active in the development of Metadynamics. It's enhanced sampling technique to accelerate Molecular Dynamics simulation, and this was really a time of big excitement in the development of a method in trying to apply it to different type of types of problems in material science, in, in computational biophysics, biochemistry. But still it was a method and we had different implementation of the methods in our research group. Uh, so there was multiple people re-implementing, uh, metadynamics in different codes, NAMD CP 2K for different types of application. And it was quite funny because. Each, uh, these people implemented their own set of, uh, collective variables, which are needed in Metadynamics. Uh, so if you want to do something with a collective variable that is not in your code, you have to copy, uh, uh, the implementation from another code. It was all very, let's say it was working, but it was quite inefficient. So we all start from there and try, at least for us, for our mental health to have just a single unified implementation of the method of a collective variable that could work with multiple codes, to make in-house the development of the method much easier. And then, of course, this naturally, uh brought us to disseminate this library to the public and to the community. So this is where it all started with the help of other people like Giovanni Bussi and Carlo Camilloni, and then Gareth Tribello, who are the core developers. And this is where it all started. My interest in, in developing codes and sharing them to the community. And then they is as, as my, as a switch from a PhD student to postdoc and especially in the US in, University of California in San Francisco with Andrej Sali. There is where we, I start getting interested in, in approaches that go beyond simple molecular dynamic simulation. So just in silico and try to incorporate different types of experimental data as an effective way to improve the force field of molecular dynamics, simulations. And the problems there was, well it's, kind of, uh, complicated to incorporate some measurements that reflects ensemble properties, not just one configuration, but the kind of full conformational landscape into simulation. And this is where we start developing uh, meta inference and other Bayesian approaches to, to be able, from average observation To determine the underlying conformational landscape. So that that has been very short description and summary of, of what I've been doing so far with my group.

Miłosz:

understand that once you have a tool, once you have a platform that makes it so easy to implement new things, it becomes very tempting Right?

Max Bonomi:

exactly, let's say what I do is always trying to exploit and make the best usage of Plumed, which is extremely flexible, and it's just a way to kind of modify your, uh, your energy function, your force field, in Molecular dynamics simulation, and this modification can be anything, can be a, a method to accelerate sampling like metadynamics or a modification to the force field to incorporate experimental data. But in the end, the, the end of the day, it's always the same thing: you want to add external forces to your mD to do something. And so Plumed is built and is done for that.

Miłosz:

Right. I find it pretty amazing. About our, let's say, simulation community, that we somehow find ways to standardize the codes and to have things that are transferable. it's not the case with the QM community, for example. They always have their codes, uh, divided into 50 different labs. And uh, it always makes me a bit reluctant to take up the codes. With coarse grained is the same. Every lab seems to have their own coarse grained method and only they know how to use it. So I think this tendency to, to have platforms that allow for cross communication between codes is, is amazing. I really appreciate your contribution there.

Max Bonomi:

Ah, thank you. I, I think we could do even better because in, in any case, there's still things that are very code specific, I think, of, uh, input files or trajectory formats that kind of, maybe could be improved in the future. And I'm sure there are people discussing the community, how we can make interoperability between MD codes, classical MD codes even better. and I think that something we can do still.

Milosz:

Right, but already with the Plumed Nest initiative where you share the inputs and, uh, well, the transparency of it all, it's, it's a great, uh, way of making things reproducible. Replicable.

Max Bonomi:

Yeah, that was something that came, not in the early days of Plumed, uh, but came at a certain point the realization that there are a lot of people out there that are using the code after all, and they have their own, uh, protocols and input files, to do specific task. And most of the time they, they are buried in a laptop or in a desktop and not shared with the community. And people can learn from these things. Maybe they want to do something very similar or be inspired and it's very difficult to read the paper and find everything, all the details that you need to redo the same thing. It's much easier to, file a link and, that's it. That's your, your method is one line. Go to this key tab, and this is the method section of my paper instead of a list of, not complete details. And so this was the idea behind Plumed Nest, just to share better what we do with Plumed

Miłosz:

right. Maybe the next step is to train a chat GPT with Plumed inputs, to make it automatic.

Max Bonomi:

I agree. My dream is to create, to learn from this and create a, a an assistance for Plumed to debug problems and in the input files just by learning from good examples and help people in an automatic way. That's my dream. Yes.

Miłosz:

That would be amazing. But, so let's shift from archeology to, to futurism so which new staff gets you the most excited about the future of the field and uh, what are the revolutions that, you know, you see them coming and you just can't wait for them to happen?

Max Bonomi:

Yeah. So it's very difficult to answer to this question because working, let's say, trying to work in close contact with experimentalists, I'm not an experimentalist, but in, my research, I try to, to incorporate and use data. I'm kind of also fascinated by what's coming from their area. But I would start with the, with the obvious that What artificial intelligence is, revealing us that can do is kind of, uh, extremely exciting. It's very difficult, I think, to understand what is the limit now of what it can do. Uh, because we have seen already we have, we have this alpha fold and, and family revolution in which at least the prediction and the creation of a model, single structural model of a biological system of proteins is there. And in many cases is is extremely accurate., but we have also seen that, maybe there's something more in these methods. We just Maybe are able to capture certain parts of a conformational landscape of proteins. Not just one single structure, not just the global minimum, but maybe they learn something more about, proteins. And we have seen modification of this method to create alternative state and so on and so forth. So I'm sure that, but, what will happen next? Most likely Is in this direction is try to, I'm not seeing replacing molecular dynamics. Maybe yes, maybe no, but try to really be able with this method to characterize better, more exhaustively, conformational, landscapes rather the single, uh, single structure. And the challenge there, of course, is not just giving a model of state A and state B. Which is part of a problem, but also understanding what is the population of, of the state A, and state B, how the system goes from A to B. So it's a way, much more complex, uh, uh, problem. But I think this is the direction of, this method where it can in, in the next, I don't know when it'll happen maybe in two years, maybe in five years, but they will be able also to provide this information. I'm pretty sure, uh, there are already first attempts either to use this AI method as a sampler, but in the end you get what molecular dynamics would give you just in a way more efficient way. That's probably is a, let's say, low hanging fruit. And we have seen some, some efforts from the community, but maybe we will be able to do it. Regardless of of MD force field, be able to learn from some type of experimental data, of other information. Also information about thermodynamics so that the, output of this method would, be not just structured, but also population and maybe Inter conversion pathway and so on and so forth. And this for me would be, would be fantastic because it's all the pieces of a puzzle. You have structure, you have thermodynamics, you have kinetics and done in a way more efficient way, possibly in a more accurate way, that molecular dynamics, simulations. And this is what is, I'm excited, mostly excited for.

Miłosz:

So, so far we've seen, uh, as you say minimum energy structures predicted very easily, but the concept of thermodynamics is surprisingly hard to learn. Right? And maybe it's not so surprising if we think about what data is out there. We don't have many data of higher energy conformers with that energy, right?

Max Bonomi:

True. No, no that is the challenge. And try to see if, uh, there will be new experiment that can provide high throughput in kind of consistent conditions, information and data that, that we can use for training. I guess we have all realized that one of the key, requirement for this method to work well is the quality and quantity of, of a data use for the training. And if the information is there, and maybe we will, uh need to, to have more of this information, to be able to have good artificial, intelligent, deep learning machine learning, whatever methods.

Miłosz:

Right, and do you see it, example, learning from actual molecule dynamic simulations where you have better access to energies, free energies. Of course you have the errors of the force field, but perhaps you can somehow leverage both.

Max Bonomi:

That I think that could be already a big achievement to be able to do mD without doing md. So have a sample of, a confirmational landscape consistent with, uh, molecular mechanics, force field using my laptop in one hour instead of using supercomputers. And of course, the limit is always the underlying force field, but we know that there are methods, and I contributed that many other people's contributed to field to refine this conformational ensemble using experimental data. So that could be the first, goal, the first milestone is, is getting there. Then we have replaced a tool like md at least for kind of thermodynamics, providing thermodynamics characterization. We have replaced it with a way more efficient tool, but with the same accuracy of md plus you incorporate data and you refine the, the thing. the second, third milestone would be what? Doing from the beginning. Better than molecular dynamics, but probably this will require additional data to train, these, uh, models. It's along the way, but probably is, is a bit farther, uh, away.

Miłosz:

Yes. I think we have to recognize and really think hard of the strength of each field, right. The, uh, experimental field in giving us the grand truth, the mD field in, let's say, us the energetics and, um, higher energy structures with thermodynamics, and then AI being able to generalize from

Max Bonomi:

Exactly.

Miłosz:

That would be amazing if, we can see more, research, more great tools coming out from that.

Max Bonomi:

I agree. I agree. And If I if I, can add the one comment, then there is all the part on the of actually structural biology experiments again, there are other revolution happening there that I see from, uh, I'm not inside, let's say, but it affects my life and my research. And then it's clear where we are going with a higher resolution, in better environment in cell with, tomography. It's amazing, uh, how we can look at things really where it matters, which is in the cells and, and improving resolution every year. So it's what cryo em revolution has been in the past, uh, I dunno, five, 10 years now. We are, we are observing and we led that in the cell. And that's this for me, that I work is in strict collaboration with experimentalists is as exciting as artificial intelligence.

Miłosz:

Yeah, that's great to hear.'cause that was also one of the messages from Pilar who said that,

Max Bonomi:

Ah,

Miłosz:

also thinking, and working in the direction of helping people from the in-situ cryo-EM community So, yeah, seems you're converging on this,

Max Bonomi:

yeah, it's, it's, it's super exciting.

Miłosz:

that's great

Max Bonomi:

Yeah. Yep.

Miłosz:

a clear direction there, uh, where people are excited. And moving on to the next part,

Max Bonomi:

Yep.

Miłosz:

you see, or what do you think is the thing that we're not doing well enough as a community? What... where could we put more stress in our scientific lives, in our, values, practices?

Max Bonomi:

Uh, this is complicated to answer to this. A again, there are many things in very different domains, more technical and more related to personal qualities. Let's see. There is a, I think, a whole class of, uh, activities that we do for the community. I would say almost services to the community, but I feel they're not, uh, as well regard this as other things because they are not as measurable as the number of papers or the H index or whatever, all these, all the style metrics that we use to evaluate researchers and all the soft thing that we do to, I don't know, developing tutorials for our software or, or teaching in in schools. Uh, and the software development itself, which is very collaborative. No, we are now Take the case of Plumed. We are four core developers, but we have, uh, important contribution for the community. And so these sometimes are written by a PhD student, which is not fully recognized for his contribution to open source software. Maybe they are preparing tutorials, uh, to put online. And this is, fundamental for a computational, uh, biologist, and chemist. And, it, it's what makes The life of all the community much easier. And I think it should be valued much more when evaluating, uh, things and, and, thinking of a researcher set of skills. And we are, we are trying to, to do our best in this, in this field. And, and in doing as many outreach as possible because we believe in it and we do it without thinking, uh, of a reward. because we have fun also among us. But I think there should be a much clearer reward in this type of services to the community, especially for early career researchers.

Miłosz:

Right. One thing I think is great about the scientific community is that we have so many people who are, you know, intrinsically driven, so people just do things because they love it. But as you say,

Max Bonomi:

Yeah.

Miłosz:

benefits, it often benefits the people at the top, while the people at the bottom are working without much recognition. So i, also

Max Bonomi:

I,

Miłosz:

that,

Max Bonomi:

I agree.

Miłosz:

I also see that, people are now publishing, uh, tutorials, publishing many accessory tools. So maybe this is improving, even if is published as a preprint, but is citable.

Max Bonomi:

I agree. But But we should find a way to really make these, these contribution recognizable as an article can be, as an invited presentation to a conference can be. So we should start opening our mind that is, uh, the product of what we do is not just, uh, papers and talks and whatever, and much larger our activities and all of these should be valued, uh, almost equally.

Miłosz:

Right. I think that the typical problem with that is that There is this famous paradox where you put a numerical measure on something, it stops being a good measure of productivity, right? I don't remember the name of the paradox then.

Max Bonomi:

me neither.

Miłosz:

It was a famous thing that whenever you start measuring someone's contribution by a deterministic number, people are just Finding ways to abuse or maximize this number without contributing the work. So I dunno if you should have another AI social score for scientists or

Max Bonomi:

I, I don't know what is the solution to the, to this problem. We are kind of, uh, attached to this metrics for, kind of simplicity to see a number and judge based on the number, but I think we should move on and we are probably moving on already from this old way of doing things.

Milosz:

One thing I'm thinking of is that science used to be perhaps much more about the community back in the day. And these days we have so many new scientists, which is a great thing, it also makes science perhaps much more anonymous or you know, disconnected from smaller communities. So it's harder to know who does what. You have to somehow rely on proxies, right?

Max Bonomi:

True. No, no. Yeah. I I see what you But overall is a, a good thing. No, but we have a, a large bodies of of scientists. Contributing because there are the way to, do these collaborative efforts, there are the platforms, the software, the, the infrastructure to be collaborative and this open up contributions, of course. So this I think is my the main point.

Miłosz:

Yeah. I've seen efforts to kind of remove the gatekeeping of science to make it, uh, to make a point that? scientists are, you know, computational scientists, but also business scientists, technical scientists, and, um public outreach scientists And we have many branches which have very different, structures of rewards. Very different. Communities.

Max Bonomi:

It is true, but in the, in in the end, many of us do a little bit of everything. Um, so so it's very difficult to to have a label. You do outreach and that's it. Of course, there are people doing most that, or in a professional way, more professional way. Uh, but sometimes p group leaders or research in general, uh, I need to do a little bit of everything in their limited time available.

Miłosz:

Right. I think it's fine to have activities that don't. necessarily increase your chances of getting the grant right, but make you more visible in the community, make more connections, and in the end, make the field a better place.

Max Bonomi:

I agree. But I think sometimes I'm thinking, as you said, for. People who are starting doing this job, and maybe they don't have a permanent position yet, and so it's especially important for them if they dedicate their time, to these type of activities to be fully recognized, because it's very important.

Miłosz:

I agree. There were some steps even to recognize, for example, reviewing, right? Now if you review for big journals, you get recognition from Publons or any other platform. So I think that's a great step in that direction.

Max Bonomi:

I agree yeah. Sorry, that to to specify one thing. The problem sometimes there is that this Could be very important at, at the early stage of careers to have these recognized. But you see sometimes the editor said, no, this is too young and it is just a postdoc. Let's say, let's assign the review to somebody else. So we should also change a little bit, this scheme that I want Super famous, uh, pI, uh, to review this paper in, maybe four months. And, uh, he doesn't care about the recognition of this team because he has other recognition. So we should change also this, this schema and, make it a bit more horizontal. The review process across career stages.

Miłosz:

Yes, maybe we could start with actually identifying, all those things, all those fields in which people contribute and have some sort of, as we have now recognition for We could have also recognition for all the other fields, and then, you next step would be probably to make it a consideration when grants or any other evaluation. Right. To, to broaden it up

Max Bonomi:

Sure.

Miłosz:

include more activities. That's a great, that's a great idea. And I think we should be going there.

Max Bonomi:

I agree 100%.

Miłosz:

Okay. So if that's it, if there's nothing else to say, then thank you. so much for taking the time

Max Bonomi:

Ah,

Miłosz:

Thank you for great ideas.

Max Bonomi:

it was a pleasure to, to discuss with you.

Miłosz:

Yeah. Wonderful. Thanks so much. Bonomi.

Max Bonomi:

And I, I hope you have the best for this podcast series. It's a, it's a great idea.

Miłosz:

Thank you. uh,

Max Bonomi:

Yes.

Miłosz:

will keep inviting great scientists like you,

Max Bonomi:

Oh, thanks. That's too much. Thank you, thank you. very much.

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