Episode 32 - Robert Best, Sonya Hanson, Xuhui Huang: Driving the Future of Biophysics. What’s Next in Theory and Computation?

Transcript

Milosz 0:01

Welcome to Phase Space Invaders. Today we're wrapping up season four with episode number 32, and since it's biophysics week at the Biophysical Society, we're doing something special with the help of Pilar Cossio,. My first guest here who's chairing the Theory and Computation Subgroup at the BPS. This year, we've assembled a panel of experts to discuss the future of biophysics. We have Robert Best. Chief of the computational biophysics section at the National Institutes of Health. We have Sonya Hanson leading the structural and molecular biophysics group at the Flat Iron Institute, and we have Shui Huang chairing Theoretical Chemistry at the University of Wisconsin Madison. Three distinguished guests who all work close to the atomistic level, but look at it from very different angles. Robert has been deeply involved in force fields refinement and protein folding. Sonya is at the forefront of augmenting the atomistic interpretation and analysis of cryo EM data, and Xuhui develops advanced models of molecular kinetics with the use of machine learning and AI techniques. We're touching on many burning questions. Our field is facing from the balance between physics and data in AI training and deployment to education of young scientists To which experimental techniques might bring about next breakthroughs in theory and computation, we don't have simple answers to most of them, but if we left with more questions than we came with, that's a sign of successful discussion to me. As per legal requirement, I should point out that the views presented by my guest. In this panel are their own and not those of their respective institutions. I do hope we don't say anything that could get my guests into trouble, but we're definitely sharing a lot of opinions here. And, uh, if you know anything about scientists, some are opinionated. I hope our conversation not only gets you thinking, but also inspires new initiatives that, uh, the community sees as crucial for moving forward. But if you're just tuning in to chill and enjoy a bunch of scientists discussing the future of their field, you know, who am I to tell you what to do? Let's go. Xuhui, Sonya and Robert, welcome to the panel.

Robert Best 2:32

Thank you.

Sonya Hanson 2:33

Hello. we're discussing the future

Xuhui Huang 2:36

Hello everyone.

Milosz 2:36

so we are discussing the future of computational and, theoretical biophysics here. And, um, maybe it's instructive to look at the last decade for hints of what's to come. And, you know, I think it's just kind of striking how much better we've got at connecting different scales, in the last decade, thanks to all the data that has been coming. And I mean, biology has this particularly tight coupling between temporal and, and spatial scales. And I would open the discussion by asking how is the community addressing this multi-scale challenge that biology poses for us? Do you think there are some gaps that we should more explicitly address or focus on to, you know, more fully understand life as a phenomenon of, physics? Sonja, they want to start as the person who works at maybe the largest assemblies.

Sonya Hanson 3:28

Yeah. I mean, I think right now there's not as much research as maybe there could be extending across these different skills. I think we're doing very well at simulating like atomistic molecules, like proteins. but yeah, I think we've, we now have, some work on, you know, nanometer scale, hundreds of nanometer scale, potentially microtubules. and we're, you know, even trying to go larger than that. And I think that how to bridge those scales, take the information we know from atomistic simulations to inform us on what's going on in these more massive scale. processes is super important, but I think there's a big gap there. and I think, yeah, I'm just kind of learning about like what people are doing in these meso skill simulations, but I think there's a lot of very cool stuff that could benefit a lot from like our atomistic understanding of things. And I think it is exciting and I think there's a lot that can be done there. Uh,

Milosz 4:27

Right. I guess we are all close to the optimistic level, right? And we probably learn to extrapolate from that to, to the actual biologic biomedicine. A lot of us probably have connections to medicine and, and human health, whatnot. But, um, where are we here? Robert.

Robert Best 4:43

Yeah, I mean, you kind of brought up something I'm very interested in at the moment, Sonya, which was, yeah, I mean, really using atomistic simulations to help parameter as, uh, you know, more core screen models and it's a really, underutilized source of information. I mean, um, I think Mo actually the most successful course grain models, I'm probably gonna get into trouble for saying this are, um, at the moment are the ones that are done top down just from experimental information. and you know, that's great. But of course, you are inevitably limited in how complex the model can be because there's only so much, you know, resolution you get from, from experiments. So, you know, it's something typically like, one parameter for each amino acid or something like that. And, you know, attic simulations are really a huge, trove of information on the, you know, detailed, length scales of interactions that, could be mined to make, more, complex, uh, course grain models.

Sonya Hanson 5:40

Yeah, and I think that in like existing mesoscale models, my understanding and there's a, is there's a lot of parameters that you can adjust, and it is just, it seems like you want to take as much information as you can possibly get, including from atomistic simulation, to figure out what those like, what realistic parameters really should be.

Xuhui Huang 6:00

I, I think I can probably follow up a little bit on this point. so I, I agree. I think the standard way, if you think about, people approach these, uh, mesoscale simulations, right? People are thinking about in the, in the old days, people are doing these theological theories. They're even simple analytical model to fit to experimental data. they're convenient, but they're often give not good predictions for new system, and they basically tell you nothing about the underlying molecular mechanisms. On the other hand, as you guys talk about, we have this bottom up. molecular simulation from, computational chemistry or biophysics. We can really build up a coarse graining model, but I don't think that really reach the mesoscale. one promising direction in my mind. We can think about roles, phase field models where people using chemical engineering continuum models, which can connect co cleaning model further to the larger scale macroscopic phase transitions, uh, in biology or any other cellular scale phenomenon.

Milosz 7:03

All right, so we're, struggling with transferability, right? this is going to be this kind of predictive versus data driven challenge of how much can we only explain, given the data or learn from the data versus As a community, How much can we drive research with predictive methods? How much can we be the exploratory factor? And it's, uh, we should probably, ideally do both, but, uh, certain things are only working at certain scales uh, certain things are applicable to all of them. Any thoughts on this?

Sonya Hanson 7:35

Well, one thing that makes me think of right is you know, your experiment can only really tell you so much, and I think that, you know, one of the things that's nice about coming from like a atomistic perspectives, you can learn, you know, what is the impact of adding a phosphorylation or a post-translational modification. That you can do in like a physical physics-based model and, um, understanding kind of how these physical interactions can inform your mechanism and understand kind of these more mesoscale phenomena. I think that like going back and forth between these scales, looking at what experiments tell us, looking at how, you know, starting from bottom up coarse graining can, like, really lead to these insights, I think is a, yeah, I think balancing also kind of like how to collect the experimental data in, in this slide is also,

Milosz 8:22

Right.

Sonya Hanson 8:23

part of the, the question.

Robert Best 8:26

Yeah, I think it's also, That's something that's advancing a lot, of course, is, uh, high throughput, experimental, data is getting to be like of a lot higher quality. And, that can also be useful for parameterizing, uh, mesoscale models. just'cause you have once you, let's say, scan across a huge space of proteins and or whatever b you know, that's, that's a lot of data you could potentially use to improve a coarse grained model. Um, of course you could also use it to just train the machine learning, uh, and AI model directly and cut out the simulations from the loop entirely, but it can also be used to improve simulations and hopefully get a mechanistic understanding of what's going on.

Xuhui Huang 9:11

I actually have a question on this point. What do you guys think? About this, uh, data dreaming versus predictive models. Uh, do we really need physics? Uh, if we get enough data, then we can probably build a perfect model. We don't need any physics or do you think physics still are important?

Sonya Hanson 9:32

I think you at least need like physics to understand the data based model. I mean, but I think it is, yeah, it is really tempting to be like, okay, if we collect enough experiments, if we have enough data, then like we don't need all of this other stuff. But I think that in biology it's so, there's so many things that go into any given phenomena that there's a lot of data there that needs to be collected. So.

Milosz 9:58

Yeah, I guess alpha fall got much better just by assuming that the polymer has to be continuous. Right? Maybe that's not very advanced physics, but it's some physics.

Xuhui Huang 10:07

Yeah, I agree. I think still the combination of data driven and physics based models still are powerful'cause we really don't have unlimit data.

Milosz 10:15

I think, well, I,

Robert Best 10:17

I mean the more physics that you, oh, sorry.

Milosz 10:19

I can chime in. I, I think it's, it's a very important question to ask. should be the prior, what should be the, machine learned part, right? cause you can, you can set this question in many different ways. what I was, I was going to ask the question, you know, what, what are we learning from the fact that bottom up cross green models. Don't really work that well. Right. as Robert alluded to before. does it give us any, any hints at maybe we're getting something wrong or are we just trying to make things too simple and they're really not simple. Okay.

Robert Best 10:48

I think, well, the problem is generally speaking that they don't generalize, because people maybe, uh, train them on, you know, train a very sophisticated model on limited data, and so there's a sort of a certain amount of overfitting, and then when you go to a different system, it's like. Your performance is terrible. But I think that if one does things in a careful way, it should be possible to, to avoid that kind of problem if you're not just, if you're, if you're actively looking at making it, generalizable to, to other problems. And, cause the thing about, you know, the sort of standard learning methods that are out there for training coarse grained models is they're very good at learning and they'll use all the available parameters to, to fit what's there in the, in all atom model. But there's of course, the, the issue of overfitting then arises. So you have to just take appropriate care to avoid, uh, that happening.

Milosz 11:41

right, so regularization matters.

Robert Best 11:44

regularization.

Xuhui Huang 11:45

I think that's, uh, Robert, Robert brought up a very good point. Uh, I'm not very expert of this topic, but I was also wondering, uh, the, currently the bottom up cross screening can go up to the modeling of the entire virus, right? But if you want to even go to larger systems, is there enough theoretical framework, like the, can you go to a continuous model? this face, field based models, uh, I'm sure people have thought about this, and Robert, maybe you can comment, have people tried under the field and what's the problem there?

Robert Best 12:17

uh, using a, as a continuum based model for, for much larger scales. Um,

Xuhui Huang 12:22

then you're taking parameters from bottom up. Code screening. I,

Robert Best 12:27

I'm actually not, that familiar. I mean, you know, I mean, the.

Milosz 12:32

Yeah, we had Z Schu on the podcast, uh, just two episodes ago. So she's, uh, doing a lot of work in that direction, which is modeling of an entire cell with, a mixture of

Robert Best 12:44

Oh, right. Yeah.

Milosz 12:45

Um, equations, like master equations and actual structural, diffusion and so on.

Robert Best 12:51

So that was, yeah, that was really like a tour force. Yeah. That was like a tour force of, you know, putting together a lot of really, complex, uh, computational methods Yeah.

Milosz 13:05

yeah, so we

Robert Best 13:05

um.

Milosz 13:06

stuck at the, level of choosing what the AI should learn, right? Uh, what the loss function should be, what the regularization should be, what the physical constraint or physical form, might beat it to get it right and to make it generalizable. Am I getting it right or

Sonya Hanson 13:22

I think often it just depends on what is the question that you wanna ask. Right. I think we're not quite at the stage of making things fully generalizable. I think there's still these questions of like, you know, are you interested in modeling like cell division? Are you interested in modeling other processes where, you know, individual molecules can be very important. and I think understanding what are the important things to capture there, I think is still maybe a little bit of an art. Maybe eventually we'll learn enough about it that it can be generalizable in the future.

Milosz 13:54

work for humans in the age of machines. But getting to that, yeah, we are probably at the stage where the kind of initial cloud of dust is settling around ai, right? And different, flavors of ai. Are we getting to the moment where you can kind of start to evaluate, you know, what will stick.

Sonya Hanson 14:14

Okay.

Milosz 14:14

And what will fail? What are the challenges and opportunities for computational biophysics research, coming from ai and how does it also differ across different career stages? Right. How does it differ from someone who wants to learn biophysics to someone who just wants to study and to, to do research in biophysics?

Xuhui Huang 14:32

Uh, yeah. I'll start with a little bit of the, what I think, I think we, we sort of touched upon this issue in the previous topic where the, uh, predictive model versus data driven model. I think the real opportunity here for power physicists is not AI replacing physics, but AI working together with physics based models. so we, uh, came from the molecular simulation field. We know in the simulation side, we already have examples of combining AI with the physics based models, right? We can accelerate sampling, learning connected variables and do a data analysis. Uh, so I, I think that that is a way, to move forward these combining physics with, uh, ai.

Milosz 15:17

Right. That's what we would love to see, instead of just using AI priors as as physics, right In the end. But

Robert Best 15:26

Right.

Milosz 15:26

there? Like what are the pathways to, to that?

Sonya Hanson 15:28

I think a lot of it goes back again to kind of what is the data available, right? So like the success of Alpha Fold is heavily based on, you know, the PDB existing. And that took a long time to collect all of those structures. and then I think also another thing important to consider about alpha Fold is that, you there was CASP which had very strong metrics of telling, you know, deep mind if they were performing well or not at this task. And so I think those two things, the data and the metrics are, you know, critical to understanding what's next. And I think, you know, at least from my perspective, one thing where like we don't really have the data yet is on like, you know, protein, confirmational, heterogeneity dynamics, kinetics. Like this is something where like we don't have something like alpha fold. And I think, uh. It's actually such a diverse kind of data set. It's hard to even think about what that would might look like. so to me, that's kind of the future is connected to all these different

Milosz 16:29

So should we strive to organize more CASP like competitions in different directions?

Xuhui Huang 16:34

I agree with that point. Uh, I, I think, uh, we need data set, which were shared, standardized, and community maintained. I think maybe there's opportunity for molecular dynamic simulation community where we have been talking about, right. So the putting together this MD simulation data set As a standard community, maintain the dataset so we can train machine learning models for protein dynamics. yeah, that would be something I think I, I would very much love to say in the near future

Milosz 17:04

Yeah, this is happening now. So there are now repositories that are starting to trajectories just like the PDB treats structures. And I think it's exciting. We, we still have to come up with very strong ideas for what you can get from dynamics that you cannot get from structure alone. Uh, especially if they're short dynamics, I guess.

Xuhui Huang 17:24

What bio email, So I think the, the current active topic will be the community trying to move one step further from to predict structure ensemble. It's not dynamics yet, but those short MDs simulation will provide abundant information for us to predict ensemble of the protein structures. Whereas than static ones. I think there are a number of examples bio EMU from Microsoft Research.

Milosz 17:51

Yeah, I know Sonia, maybe that's, uh, overlapping with your expertise there. Cryo EM data could give us some insights into you. Those subtle proportions or subtle ratios of, different confirmations coming directly from cryo EM data, right?

Sonya Hanson 18:07

Yeah, this is definitely something that, you know, cry, cryo eem. more and more has the capacity to do is give us these ensembles from an experimental technique. but I think that like cryo eem isn't the only experimental method, um, that can give us some of this information about these ensembles. and I think that cryo is still quite labor intensive. And so I think thinking about what are like high throughput experimental techniques and can really give us. Information we would need to understand better, I think is is worth thinking about carefully.

Milosz 18:43

Yeah, it's pretty tough because every case where you have a dynamic heterogeneity of structure, you can have so many labels. If you wanna use fret, you have to interpret it. You, if you wanna do free energy methods, you have to interpret them. You have to assess convergence, right? So like a lot of those things are very case by case. And I wonder if, if that's kind of amenable to this machine learning friendly workflows where you have, you know, clean data, same format, um, as sort of repo where everything can be standardized.'cause this looks very tough to me.

Sonya Hanson 19:16

I think if you, if you can map these observables back to like. What do they mean to your protein structure then it can be easier.'cause then you're always mapping it back to the same thing. I think even in cryo-EM we have this issue that often, you know, you get out these volumes that you still have to fit the structure into. so I think there's, every experimental method has like the its own thing that it gives you. You need to figure out how to connect those all together. tricky.

Robert Best 19:44

I mean, I guess it's also what, you know, what timescales, I mean. I mean, there's just, you know, fluctuations about the native state, I would like to say are relatively trivial. You could do an elastic network model and, um, save yourself a lot of trouble. but I mean, if you, you know, if you're looking at more profound like rearrangements that might occur on a, you know, millisecond timescale, I mean, you're not gonna see those and you may not even guess that they would occur in a just unbiased MD simulation. So that's where it's kind of, it's hard to, just do a, a large scale, simulation effort unless you're able to really run, I mean, now I guess maybe in the next few years, millisecond timescales will become more commonplace, so maybe that is gonna be a possibility. of course, then you have to worry about the force field and everything.

Milosz 20:32

is very true that what you can get from a simulation is very contingent on, uh, what is available, what kind of movement is available within a given

Robert Best 20:39

Yeah,

Milosz 20:39

right? So like, that's part of

Robert Best 20:40

I mean,

Milosz 20:41

that we're going to get something, but there's no uniformity of. heterogeneity that we're generating, right? So like some things will be completely folded for the entire simulations And some things will actually sample some relevant unfolded fold. The dynamics so this is going to be another big challenge, how to, how to disentangle those.

Robert Best 21:00

And as soon as you go away from the folder state, then you sort of worry more about, uh, you know, how much hope, how much can you trust the force field, uh, in, in doing that. Yeah.

Milosz 21:11

That's another worry. But

Xuhui Huang 21:13

Can I add one point, before we move forward? I was thinking that now we talk about protein dynamics. That's obvious to the next topic. Uh, people are working to develop machine learning model for protein dynamics more broadly. If we think about rose physics based, uh, foundation models. Now, when we talk about larger language models, the protein large language model, turn on evolution sequence are very useful. And you just need to do a little bit of active learning based on the protein, large language model. You can predict a lot of experimental data. Different task. Can we, can our community build physics based larger language models, right? We can use molecular dynamic simulation as training data. We can use. Other thermodynamic example, free energy of mutations, for example, solvation, we can build large language model based on physical principles. I think that would be a very, a very interesting direction to go because we're not limited by evolution signal anymore if we do that

Milosz 22:14

So you mean, uh, a model that will be trained on predicting various physical properties the way that large language models predict. I don't know, missing letter or missing token.

Xuhui Huang 22:23

right. So it's more like learning or embedding. So you need into the sequence, you output the embedding, but that embedding is learned based on the physical, based, uh, dataset. Could be molecular dynamics simulation, could be, some other physical properties or thermodynamic properties. Uh, so then that will be non limited for the, uh, evolution or that can be combined with evolutionary data.

Milosz 22:46

That sounds very challenging, but interesting.

Sonya Hanson 22:49

Yeah, I think there's this question of how much is encoded in the evolutionary data like. Is there dynamics in there? You would think that if it's important for function, it should, should be some signal, right? But how much can you really learn from it is another question.

Milosz 23:04

Yeah, there was this idea of dynamics floated around a while back, right? So like different signatures that might have different functional meanings how different parts of molecules correlate and so on. So probably that's one idea

Xuhui Huang 23:17

I, I think the Sonia's Sonia's point is a very value. Recent studies show the, a lot of the protein dynamics are related to functional already encoded in the evolutionary signal. so if you want to build. The foundation models, I think the better way will be combining those dataset from the physics, from the evolution, that would be more powerful in general,

Milosz 23:39

Yeah, it will be pretty, pretty exciting even to predict like flexibility from sequence, right? Things like that or

Xuhui Huang 23:45

or predict solubility from sequence, something like that.

Milosz 23:48

I think there are, there were already models doing, trying to do that protein suitability enhancers based on language models, but uh, for sure there's gonna be a lot of work into that direction.'cause it's so also commercially interesting, right? Could save experimenters, a lot of work, but moving a bit beyond, Physical models. what kind of challenges? the, let's say cha GBT revolution, is, putting in front of

Robert Best 24:13

yeah, I think it's, uh, certainly I, you know, I, I have to first of all preface this by saying that I don't use chart GPT or I've only used it very, very level. but I can see that, that this sort of, LLM base. Uh, AI and maybe other kinds of things that may come up in the future. could be very, uh, powerful tools. Like, you know, not like directly in our research, like the, what we've just been discussing, but indirectly. Um, for example, in writing code. so I guess a problem, you know, we have the opposite problem of software companies that have too many coders. We have like too few coders generally speaking in academic, academic research. And so if it was possible, you know, if AI is able to help, writing code, then it could mean, we may make it more feasible to make, you know, larger, more robust, academic software packages rather than, you know, what we have now, which is held together with some chewing gum and stuff. Um, That. So that's one thing I can definitely see, but yeah, I'm just seeing, I, you know, I, I wouldn't, first of all at the moment, trust any, code that, was produced by, um, AI without checking it myself. But, you know, it seems to be very powerful at, at doing that. the other thing is it seems to be very good at doing math, like amazingly good. and so I, I, I suppose that you could imagine in the future, people might use it as a sort of a tool, like a, you know, might currently use and then of course there's writing, which is, uh, maybe whatever, thinking about like, uh, writing papers, writing grants and so on. And, I'm very averse to using it for those sort of things because I like to, something I enjoy doing is writing the paper myself, but I, I suppose that, you know, people are, are doing using it more and more. I.

Xuhui Huang 26:11

I'm more optimistic and, uh, about this point. so, uh, I was teaching the, uh, undergraduate machine learning in chemistry class, uh, in my department. And I, I have my homework, which are coding homework I to students, and they, many of them try to use charge GPT to write the code. so unfortunately the successful rate, without any prompting or fine tuning. It's very low. It's like 20%. You always have box, I think the way to go would be this AI agent where you couple the large language model with a reasoning agent. Execution agent, where you can automate as Robert mentioned, you can automate part of your workflow, right? Like we do MD simulations, like we're doing this data analysis. We build a Markov state model, uh, using this AI agent. That, that's something I think my group started to work on, and I'm, excited about that myself. And of course, there's always a validation step. You have to make sure the results are correct.

Milosz 27:10

Yeah. Is there a danger that we're gonna skip the validation altogether when we are, when we get used to the fact that it mostly gets it right? Right. eventually if it's like 95% correct, we're going to say, oh, it's basically 100% correct.

Xuhui Huang 27:25

Right. I think there's always two camps. One camp of people think the larger language model are better to serve as a code pilot, right? We help you write the code, just provide input, but you still write the code, validate yourself. Another campus is want to everything to be automated, right? You have the close loop. You don't even have human in the loop. You just pile together like 10 different agent. Then you can, you can start doing things that can generate, uh, design of proteins or antibodies. Uh, so yeah. So still I think there's the debate between the two camps.

Milosz 27:57

Definitely relating to what, Robert said. I think there's a great case for, debugging and maybe documenting scientific software, which is something that doesn't necessarily write the code itself, but can signal a lot of issues and, and Improve the experience of interacting with the code because, um, you know, people can still review what is, what is created and make sure that it makes sense. But there's so much

Robert Best 28:21

Yeah.

Milosz 28:21

or poorly written code in, in academia, and I've certainly contributed my share to that. that maybe this is going to be like deploying these agents on academic, um, let's say repositories without letting them modify things, but just signaling things, making, you know, pull requests, whatever

Robert Best 28:40

Yeah. I'll be just writing an API documentation I think would be something that it could do very well. Um, and yeah, writing more user friendly, uh, documentation. Yeah, probably also be done. Um,

Milosz 28:53

People also talk about writing

Robert Best 28:54

yeah.

Milosz 28:55

friendly documentation now, right? So that someone who asks how can they use this software can actually get advice from GPT. Uh, on using this software. Sonia, any experiences on your side, uh, in that regard?

Sonya Hanson 29:09

I think I have like somewhat similar experiences, like yeah, using it to help a bit with code. I'm also curious if, if it's gonna be used to, like, there's certain codes that are quite old and like maybe, uh, you know, as hardware improves, like the codes themselves like don't really improve. So, uh, and maybe like the developers have like gone onto something else and they still barely work, but are still maybe heavily used just'cause they do work. So I think it'll be interesting to see kind of if there's some sort of, growing usability of existing code there. that yeah, often is maybe written by a graduate student that moved on, but is still used by the community for,'cause it works.

Milosz 29:54

Portability,

Sonya Hanson 29:55

but I think also,

Milosz 29:57

Those things.

Sonya Hanson 29:59

exactly updating and just looking at simple GitHub issues that, you know, uh, an AI could answer. it will be interesting to see how it's used. I don't know if academia is maybe, uh, as up to date as industry and those sorts of things or, I don't have a good sense of like how industries maybe using AI for that kind of thing. but I also, I don't know, I find myself, while it's like there's a reason to be optimistic for how, you know, AI can help us with like writing codes or, you know, agents to do something like, building Markov state model, I think. You know, you mentioned earlier kind of the different generations and different experiences with this. I think there is a concern of like education and like, for us we kind of went through and had to do all these things by hand. And so we know very viscerally like what can go wrong and what needs to happen correctly. But I think for, you know, students and postdocs who like don't develop that experience, um, is there something that's gonna be now like not there in the training that could eventually lead to problems and mean like that the people in the next generation of group leaders, like, you know, it's hard for them to troubleshoot certain problems'cause they're just used to the AI solving all their problems for them. I don't know. So I think there's big questions there and I think maybe Xuhui also can comment. Like, I find myself very worried when it comes to like how, um, how people are developing as writers. Because I think a lot of people useche PT as like a tool to help with writing, but I think like the process of editing and going through things over and over again manually is also like a big learning

Milosz 31:40

I structuring your thoughts

Sonya Hanson 31:41

I think will also be,

Xuhui Huang 31:42

Right.

Sonya Hanson 31:43

yeah.

Xuhui Huang 31:44

I, I agree. I think the, uh, in terms of education, that's a big problem. Uh, in general, I think graduate student or postdocs, they, they develop intuition, right? By doing many low-level tasks like writing code, running simulation, processing data, that's very important for them to formulate, to get familiar with what's going on in the field. But if this all replaced by ai, then that, that could be an issue. I think, I would still keep at least some of this component in my own lab, least to help train the students, yeah, in parallel with, ai. Uh, so yeah, I think in the future, another thing I'm thinking about in the future, what we need, right? So maybe the most valuable skills may from the running simulations, writing code, execution task. Thinking, right? Thinking clearly about the problem. Formulate the right questions, design the approach or critically evaluating the results, right? I don't know. So this, this abilities relies more on the logic, reasoning, and conceptual understanding rather than technique execution. there's always a balance. So it's still a hard to tell at this moment. That's my feeling.

Milosz 32:54

True, I think. I think the biggest challenge for me when I use AI tools is losing this kind of situational awareness, this mental model of what the tool that you created is exactly doing under the hood, right? Because like when you created it, you know how it operates and when you didn't create it yourself is really, really hard to figure out what's going wrong when it's going wrong, or. Uh, what could even go wrong it's like this mental model of what you're working with is going to be important to retain. That's my, guess here. there are for sure ways to do that, but they're not going to be automatic people just fully rely on, on AI there know that matches your, your experience.

Robert Best 33:34

It's kind of, it seems like it's gonna be a general issue, like for all AI that, where it's replacing some task we did before, like. are we gonna maintain that expertise or Well, our brains just turn to Jay or something.

Milosz 33:49

That's going to be tough. to,

Xuhui Huang 33:52

Well, I, I would think that's not only applied to computation about physics, but also applied to many other fields. Right. That's a

Robert Best 33:59

Yeah, yeah. No,

Xuhui Huang 34:00

society. Yeah.

Robert Best 34:01

absolutely. Yeah. Yeah.

Milosz 34:02

right, so I, I wanted to stay, uh, a little bit on the topic of education of young researchers. Do you think now that we are kind of having a broader view, thanks to all the data, all the, all the insights, are we really, creating, you know. New researchers who have a bird eyes view of the, of the field. Should we actually do that? Maybe, should we give people a broader view of biology, a kind of more, as we already discussed, multiscale or I integrative view of, biophysics? Or is it kind of fine for people to go deep into every, like corner, uh, defined by every new method that is, you know, popping up every once in a while? Like every time there's a new tool, people being experts on AI prompting or alpha follow three and so on. how is this balance going to evolve in your opinion?

Sonya Hanson 34:53

I guess I do feel that more accessible now for like a PhD student to be able to do both computational and experiment. like in Cry Em and Crystallography for example, there's Model Angelo, which helps a lot with model building, which used to be incredibly time consuming and now, you know, it's still something you have to monitor manually, but at least it's cut the time down a lot there and then opens up, you know, for students to learn maybe some more computational methods during that same time period. I think it's very interesting and I also think there's like a lot more like experimentalists, I think respect computational methods a lot more now. I don't know if you guys have this same experience. Well, yeah, being the artwork.

Xuhui Huang 35:37

than 20.

Robert Best 35:40

Yeah, well, the methods are better than 20 years ago. So, yeah, yeah,

Milosz 35:47

Yeah, people are definitely coming more to ask for help with something that they want to want to figure out. also my

Robert Best 35:53

yeah. I, I definitely think that, yeah, that, I mean, compared to when I was a student, I mean, yeah, there are more students doing, uh, interdisciplinary work, whether it's using, you know, multiple computational techniques or maybe doing experiments as well. So, I mean, I don't know what the students themselves would say, but I think that, um, you know, I think that it is, uh, it is more interdisciplinary. Used to be.

Xuhui Huang 36:19

I, I think the, from the education perspective, I think the next generation of computational biophysicist have to be multilingual. Uh, at least I would think about three pillars. The first one will be the statistical mechanics, physics theory. The second one is the data science machine learning. The third one is biology. I would very much hope the student, uh, have exposure to all these three, uh, component before they dive into a specific field, like you mentioned, maybe doing molecular dynamics, doing the Markov state. Uh, so then they get a comprehensive, training and they, they form their own way to think about things, and that'll be paved the way to become independent researcher. But that, that may require student taking a slightly larger number of the graduate courses, and participating in some other, uh, training workshops, so on and so.

Milosz 37:12

Yeah, I was always thinking, you know, how to integrate the person who is, for example, a great specialist in AI but has no idea about physics or biology, right. In a computational biophysicist group.'cause there has to be this shared vocabulary at least, For people to be able to

Xuhui Huang 37:28

Yeah, I, I want to comment on that. That's a very good point. I think I want to just spin this other way around then, if we have, the biologists know nothing about machine learning, how do they get into the field? So I was, when while I developing this machine learning chemistry class, I realized a lot of the biologists, if they want to take computer science courses for machine learning, many cases, many times they get confused about the mathematic background and the theories. And they probably want a course tailored for biology or for chemist, for learning machine learning. They want to have conceptual understanding. Some practical experience of the, for example, building up your own coevolutionary neural network. Recurring neural network. So this kind of, uh, course training coursework. I think it's also important in the future we need to develop courses targeting, uh, people from different field. That's, that's my own take

Milosz 38:24

All right.'cause everything is available with language models, for example, but not everything is in your mental latent space, so to say,

Xuhui Huang 38:32

in your language. Yeah.

Milosz 38:33

yes.

Robert Best 38:33

But there's only, there's always room for people who want to go very narrowly into, you know, very deeply into a certain area. I mean, you know, as long as they talk to other people that that know more about other things, then you can be some benefit for everyone.

Milosz 38:49

Yeah, there's definitely something to be said in, favor of deep expertise, but as you say, very often it might go too deep into like not communicating with the rest of the world, and that's the thing to avoid.

Robert Best 39:01

Yeah.

Sonya Hanson 39:02

I think, I think definitely like different fields can have different languages sometimes, I think obviously deep research is good, but I think sometimes, you know, if you're somebody who can kind of speak two languages of two different fields, there's like a big opportunity too. Kind of find, find low hanging fruit in between them and really do a lot of really cool stuff. I think it'll be interesting to see if like more interdisciplinary people existing will kind of generate some new cool ideas.

Milosz 39:32

Or just more conferences, who knows which way it goes? No. But, uh, I'm, I'm certainly think of myself as a person in the middle and, uh, I always appreciate the people who actually have, you know, this deep, expertise, but also can tell me something new about things. I'm happy to see, see both branches grow. But then as we were speaking about the languages developing internally or, you know, between the fields, How do you see different fields or sub fields redeveloping their own languages? I'm thinking of, what are the, approaches to generating new golden standards best practices and, teaching as, as the new methods become popular I think a lot of our standards from maybe a decade ago have to be deeply updated and, uh, are we doing a good job as a community in actually defining those new standards, those new best practices.

Xuhui Huang 40:29

Unfortunately, I don't think our field is doing a great job on I don't know if other panelist agree with me on that. so a couple of examples. I was thinking like 20 years ago at teori workshop people are talking about developing the benchmark systems for developing the free energy method, right? You have all different kind free energy but that hasn't been realized yet. in our community where I've been talking about, having this, uh, benchmark dataset for the, uh, development of the kinetic models Markov State model, like Milestoning and weighted ensemble. But seems we haven't, uh, moved forward too much along that direction either, this is something I personally think we should, spend more effort, for our field to develop.

Milosz 41:15

We had one podcast episode where we pushed for. A general benchmark for free energy methods, and that was probably two years ago and it still hasn't happened. So I, I fully

Robert Best 41:24

Hmm.

Xuhui Huang 41:25

So you have similar experience.

Milosz 41:27

I would love to do that, but I don't have the time. I think everyone else is in the same position.

Sonya Hanson 41:31

Yeah, I, I agree. And I, I also think that, um, seeing kind of the value of benchmarks to machine learning and the progress there, I think also sort of renews this, maybe momentum towards trying to develop these benchmarks for our own field so that like there can be more like concrete advancements and, metrics to judge tools by basically.

Robert Best 41:53

One really nice benchmark are these, you know, DE Shaw simulations where they really sample everything by brute force and you can just, uh, see if you can reproduce those results, uh, by other methods. So,

Sonya Hanson 42:08

Yeah, but those are, like 15 years old right now. Right.

Robert Best 42:11

what, can anybody run those simulations on their computer? No.

Xuhui Huang 42:14

Well, I, I, I think this is a very interesting point. I want to, I want to expand this a little bit, right? So we all know DeShaw has done a very good job, uh, generating this benchmark dataset. They're old, right? 15 years old now, I was thinking maybe in, at least in the United States, the system I'm familiar with, the funding system make it slightly more difficult, right? to really spend a lot of effort doing this kind of boring work, right? maybe the, the industry or non-profit organization, maybe a, should be the organize this kind of effort. That's, that's my thinking. I don't know what You guys think, Sonia, you are, in the nonprofit.

Sonya Hanson 42:51

No, we are thinking this way for sure. So I, I think it's a really a tempting way to go, especially given like resources and the, you know, what the field needs. So

Robert Best 43:01

Yeah, I mean, I just put the DE Shaw simulations out there as an example of where, you know. at least for some of those proteins, uh, they're, they're been really thoroughly sampled. and it's still not possible for most people to run those simulations on the, computational resources. So we would have to do some kind of enhanced sampling. We wanted to do that. Um.

Milosz 43:24

Yeah, I think one way in which. They're nice is that they are in this kind of relevant scale not a alanine dipeptide or 2D potential, which is always easy to sample with everything, but it touches on this, scale, which is relevant for many projects for actual biology. I think we also miss, I mean, there's a lot of research going into really big systems, right? From, spike, in the COVID era to like all the huge, assemblies from Cry Em that are coming out these days. And there's this big question of like, can we do anything about them? Or is anything that comes out just random noise that came out of the one microsecond, uh, simulation.

Robert Best 44:05

So you talking about reproducibility or

Milosz 44:08

I mean, I'm guess a lot of methods are applicable to a lot of systems but we don't really know what kind of systems they are applicable to Like where is the where end line or what is the trade off between, you know, size and the, and accuracy.

Xuhui Huang 44:23

right. I think this has been brought up, uh, in a number of conferences, right? When we talk about benchmark, benchmark doesn't mean 2D potential peptide. We need to have the benchmark for realistic systems, Which means, uh, we need to run very non simulations and enough sampling straightforward MD for complicated conformational change. But that's still something as, as Robert mentioned, it's not easy to do. We still have to utilize, uh, enhanced sampling or this, that type of. Algorithm To bridge the gap. But it would be nice if we have that kind of benchmark, right? If someone really spend a lot of resources, effort to generate benchmark for one or two of these kind of systems, then that would be very useful for the community.

Milosz 45:04

Well, I was advocating for symmetry based benchmarks where you know that the answer is zero by symmetry, and you can probably generate quite a few of those. but well, okay. Is there. That's maybe a different question for each one of you, but as I think so often the progress in computational biophysics is driven by experimental developments, and we had many cases where a relatively simple experiment could, you know, push computation quite far. Do you think they are experimental techniques or actual experiments that should be done to push the field forward quite a bit? what is the best case scenario for an experimental development that would, you know, make your job so much easier in the next five, 10 years?

Xuhui Huang 45:49

So I can start. so I, my group's working on Markov state modeling. I think the very direct validation or comparison of my data to experiments will be time resolved cryoEM. Where, where you can monitor the protein ensemble as a function of time, right? We get a structure ensemble, then we can directly compare with, uh, simulation without, so that's my wishful thinking. If there are good time resolved cryoEM dataset, I'll be

Milosz 46:16

And we're talking like millisecond time scales here. Right. Uh,

Xuhui Huang 46:20

We're talking about millisecond times. Most of the, uh, relevant biological confirmational changes occur at the millisecond timescale.

Milosz 46:27

and this is also where the technique can,

Robert Best 46:29

I then you were to worry about freezing and

Milosz 46:32

Yeah. There's, I was, I was going to

Robert Best 46:33

what happens during freezing and

Xuhui Huang 46:36

Yeah, I don't know what's the status. Maybe Sonia know more how far away we are to get a good resolved. cryoEM dataset.

Sonya Hanson 46:43

I mean there, there are several groups who do time resolve cryo. I think the issue with a lot of time resolve techniques though, is that you get very few particles so often. Um, the data analysis afterwards still quite tricky but I think technically possible, especially if you collect a lot of grids, uh, I think the other thing, there's, there's just like a lot of interesting developments recently as well in terms of like. Even going down to like tens of microseconds, timescales for time resolved cryoEM, super exciting and relevant to other biological systems. and then, one of the problems is that a lot of cryoEM methods do this like, uh, plunge vitrification and kind of how the temperature not only in the liquid ethane, but like above that in like this cold air column affects your conformational ensemble is a big question. And there's some very cool work that I saw last year and I think is now out where they actually just, inject the liquid Ethan directly on your samples. So you don't worry about that and you get like, very quick freezing. So I think there's a lot of cool stuff going on in cryo em to make time resolve cryo em more and more, like a real tool people can use with a lot of less, a lot fewer question marks around it. Uh, but it's still quite tricky to do, I think.

Robert Best 48:03

But it's not, it's not gonna be. Really time resolved. Right? It's, it's gonna be like a free energy landscape or something. Or you're talking about doing, um, the pulse chase kind of experiment where

Xuhui Huang 48:15

right, it's coupled with, you can do the rapid mixing, you can couple with, um, micro fluidics or flow, right? You can have triggering event and you can monitor the, the relaxation of the population. I'm thinking.

Sonya Hanson 48:28

yeah. So, yeah, often time resolve methods will have like, uh, microfluidics where you

Robert Best 48:32

The pump prob. Yeah.

Sonya Hanson 48:34

it's ligand and then you have longer chambers where they're interacting. Um, and that's kind of the time component, how long it's in kind of this interaction,

Milosz 48:43

methods determined

Robert Best 48:44

And why do you have to do okay.

Milosz 48:47

Right? And there are so many factors that can get in the way of interpreting kinetic data.

Sonya Hanson 48:53

Yeah. Yeah.

Milosz 48:54

we have any, any biophysical, computational ways of actually getting them into account directly?

Sonya Hanson 48:59

Well markup state models she approves.

Milosz 49:07

no. I'm also thinking of exactly this question of, okay, maybe we should model the whole setup of the cryo em, liquid et plunging, right. And, and do, I don't know, a sort continuum modeling of heat flow.

Sonya Hanson 49:21

Yeah. So we are kind of working on this right now. Uh, right now we are not doing the continuum model of heat flow, but we're using a thermostat to vitify samples in a MD system and looking at how the, um, basically the glass phase of water affects the ensemble distribution of the protein. right now, yeah. So, so that's an interesting study right now. We find that, uh, you know, there's, the ensemble isn't usually perturbed, but we can find some bounds over which it is.

Robert Best 49:53

But it could be perturbed on the, I mean, it's like what, 200 k or something where you hit a glass transition. Things can change between 300 and 200 also, and you take some, I don't know how, what the rate of cooling is, but

Milosz 50:10

yeah.

Robert Best 50:11

whether it be,

Xuhui Huang 50:12

So what do you guys think that if we don't like freezing, how about liquid phase TM with obvious way we can do time-resolved liquid phase TM resolution not as high cryoEM, but people are still working on improving that. Right?

Sonya Hanson 50:26

Yeah. I dunno. I feel like liquid phase is a, it's a little iffy at the moment.

Robert Best 50:31

what about NMR? I mean that's,

Sonya Hanson 50:33

NMR is great.

Robert Best 50:34

yeah,

Xuhui Huang 50:35

NMR is average. Right? So another type of technique we're thinking about. So average, right? NMR, single micro frat, right? We can further

Robert Best 50:41

yeah,

Xuhui Huang 50:42

the resolution. Maybe that's, but that's

Robert Best 50:44

yeah.

Xuhui Huang 50:45

direct comparison with, uh, istic simulations because you only, you have to do average.

Robert Best 50:50

Well, if you, do you time resolve single molecule FRET? Yeah. if the time resolution of, you know, either single molecule FRET or, um, single molecule pulling experiments or other single molecule experiments was improved to the sort of, you know, nanosecond timescale, let's say. And that would be a big game changer in terms of comparison with external, with the simulation. Right.

Xuhui Huang 51:11

agree.

Robert Best 51:12

but in terms of what I think is like the most likely, thing, I mean, it's just really, you know, development of, high throughput microfluidic, fluidic methods for, uh, measuring biophysical properties, you know, that, that may make a, a big impact in our field. I mean, in terms of let's say providing data to train coarse grained models, for example, or providing kinetic data for putting into whole cell models, think. I mean, that's definitely stuff that's rapidly developing at the moment, and that could make an impact in the next few years.

Milosz 51:46

What properties are you thinking of specifically?

Robert Best 51:49

Oh, oh, things like, um, enzyme kinetics for example, or, protein, protein ligand, protein protein interactions, protein nucleic acid interactions. I mean, just the ability to screen like, you know, thousands of biomolecules against thousands of other biomolecules, uh, is, is something that's becoming possible now. and to do that in a way that's actually, you know, comparable in accuracy to what you would get by doing a careful like, you know, conventional experiment in the lab,

Milosz 52:18

by kinetics.

Robert Best 52:20

available. I mean, of course. People would like to train our Ai, AI models on them. Great. somebody can do that. But I think it would also be useful for, for us, um, in trying to develop, you know, um, simulation models. Um. coarse grained notice for.

Milosz 52:36

is anyone going to complain about RNA not being, you know, crystallized or, I don't know what your opinion on that is, but that's another, piece of biology that is really challenging to model. I don't know if you have any overlap or any

Xuhui Huang 52:49

Well, for RNAI think even from the simulation perspective, I think we have a non to go, very simple thing of the, uh, mechanism, concentration driven, RNA folding very few force fields capture probably have to go to a very expensive polarizable force field, uh, like ameba, right? So that will take forever to converge. Uh, so still a very the RNA simulations. That's that's my feeling.

Milosz 53:16

Sonia is the Cry Em community catching up with with RNA

Sonya Hanson 53:21

I mean, the nice thing about cryo, like RNA in cryo EM is that, uh, there's more contrast for RNA than there is for protein. So, in some ways it makes it easier to do things like analyze an ensemble of RNA, but the fact is, is that RNAs are still very like, you know, capturing these, confirmational ensembles if you don't have something that you can kind of latch onto and do your pose estimation with in cryo EM and do your initial angular estimation, it's still very, very tricky, to do those analyses. You know, we're trying to develop methods to work towards that. Um, but I think RNA in general is, uh, fascinating, right? Like it has, you know, we talk about protein confirmational ensembles, but I think for RNA that we're not even, like, we're just starting to understand kind of what's going on there and, and structure prediction are way worse. and, you know, our, yeah, our computational models are way worse, way more expensive.

Milosz 54:22

Yeah, I worry if we even have a mental model for how RNA should behave in the sense of like, what is the temporal evolution of an RNA structure that is a generic structure, not just a folded rock solid.

Xuhui Huang 54:36

So I think the underlying energy landscape for protein folding versus RNA folding are different. I think RNA landscape is very rugged compared to the protein energy landscape, so would be even more difficult to sample.

Sonya Hanson 54:49

I also don't have a strong sense of like how much we know about like evolutionary trends in RNA, how much they tell us about, you know, like for proteins. It's been very valuable to look at like co-evolution patterns, but I don't know how signal there is in RNA. Maybe somebody else here

Xuhui Huang 55:07

So there, there is, I, I don't know.

Robert Best 55:08

there is some.

Xuhui Huang 55:09

curious about that question. There has been a number of the, uh, the sort of the, uh, genomic model, right? The large language model for the DNA has been recently published, like Alpha Genomics, right? People are looking at Robert, you are the expert probably on this, uh, DNA chromatin, right? But then the RNA is still different because you have all these, uh, splicing and post transcriptional modification. How much of the evolution signal has been preserved through that process, or what can we learn from that? That's I think the, it's very interesting question.

Robert Best 55:46

uh, I was just gonna say as far as the, the co-evolutionary information, I mean, there are some cases like specific road designs where, you know, they do fold into a specific structure. and there the co-evolutionary information, um, can tell you about the structure. But yeah, in general, because RNAs don't fold into a single structure that it's hard to all, you know, make sense of any co-evolutionary information on the, in the way it's done for proteins. So it's not gonna be a general tool like, uh, was, it's, yeah, it's a challenging problem. and, uh, we have a lot of work to do even at the, you know, we said the atomistic models are the most developed, but of course, Herne, even those are not quite there yet. So, yeah.

Xuhui Huang 56:28

Well, so the RNA folding itself is challenge but protein RNA interaction think is, uh, relatively simpler. That's my feeling. The transcription, transcription factor, interacting RNA this kind of think this field will be, we say the much development in the coming years.

Milosz 56:45

As long as we have a good reference Um, okay, let's move up a level and let's wrap up maybe the discussion with, what you see as, the biggest challenge for, for the community in the next few years or the next decade in terms of. Everything scientific. It can be, you know, publishing, it can be reproducibility, it can be, you know, software quality as we already, uh, discussed. Or maybe the impact of the field for biology. And, uh, what are your, what are feelings about this?

Xuhui Huang 57:16

so. I think you touch upon a few points. I still think the purely data driven model, will have its own limitations. I think the biophysics informed, uh, ary models where we, we, we combine data with some physics will be still powerful under the regime where we don't have infinite data. So that's still, I think our field will have impact. in the challenge, I think. So also the, as you mentioned, uh, how to interpret our model.'cause there have so much ass simulation data. Very complicated models. Right? So another thing would be, how do we make sure the software sustainability, right? We've seen cases where the, as as we talk about student left the group and the software, nobody maintain and how can we make sure the software, we have the software in the field that can be sustainable developed and also the, uh, the training, the interdisciplinary students as we talk about in the age of the aI and or chatGPT, that's a few things in my mind

Milosz 58:19

Okay, Sonia.

Sonya Hanson 58:20

Yeah, I think, uh, incorporating the new with the old is gonna be a challenge. So, like, I've heard this anecdote that the number of, high resolution structures that have been solved and deposited in the PDB has actually decreased, real experimental structures has decreased since alpha fold came out. you know, despite the fact that, you know, there's still a huge number, like 20% of structures that alpha fold does not predict very well. And so I think like, while, there's benefits to these new tools, I think not, you know, leaving our old tools behind and continuing to solve these structures that we need to solve, that do provide new information that aren't predictable is, is a very important. Uh, very important thing to kind of grapple with in this time. but I also think that, you know, as I mentioned, there's very exciting things going on in like, cryo eem just at the experimental level that I think coupled with the new tools will, potentially advance us understanding things like kinetics and confirmational ensembles much better. I think that's very exciting.

Milosz 59:23

Well, maybe everything that could be resolved to a high resolution has already been resolved. Now we're, we're doing challenging ones.

Sonya Hanson 59:31

I don't know. I don't believe that. I think there's still work to do.

Milosz 59:35

Okay. But anyway, be aware of feedback loops. What is what you're saying, right? That there might be behavioral, nudges that, that make us do unpredictable things in the future just because of the tools that we develop.

Sonya Hanson 59:47

Yeah.

Milosz 59:48

Yeah. I think it's good to have conversations around it and then highlight those dangers. Robert.

Robert Best 59:54

Oh, the biggest challenge. Yeah, I think, I think, I mean, showing that, you know, mechanistic or physics based models combined as needed with the ai. You know, is still useful. And, that's gonna, I mean that's, you know, that's our, I guess our challenge, uh, for, for the next decade. I mean, I think that they are, but I know that there's going to be, there's going to be opinions, uh, to the contrary that, you know, we don't need anything beyond just a data-driven models. So I think we need to prove ourselves, that we can provide useful. I mean, I think we can provide. Information. Um.

Milosz 1:00:30

Hopefully, maybe to wrap up, what would be your message to someone who's starting a PhD in computational biophysics this year? Let's say someone who wants to do research doesn't know where to go yet, but uh, you know, they like computation, they like bioinformatics, biophysics, and, uh,

Sonya Hanson 1:00:45

I liked Shea as a three pillars. I thought that was like very insightful that, you know, uh, let's see if I can repeat them right. It was like stat me and data science and biology, and I think like all three of those are kind of different skills that, you know, maybe any individual might, know, excel at more or less. But having kind of a solid background, all three of those I think is very, I like that idea.

Milosz 1:01:15

Yeah.

Xuhui Huang 1:01:15

Yeah, so I, I think, um, Think that that will be a good idea if the student entering this field can spend some time the exposure to the three pillars, then that will help them to make the informed decisions what they're gonna do in the future. So I think the many universities have this rotation programs and the first year spend more time to expose to different aspect. And then at the end you can make a decision on which field you're gonna get into. don't rush. Uh, take your time. Be patient.

Milosz 1:01:45

Great. Thanks, Robert.

Robert Best 1:01:47

I'm trying to think about what are the important problems, uh, in the field, I guess. I mean, that's not something a student isn't gonna know at the beginning, but as they go through their PhD, they can develop the opinion, but to think about that,

Milosz 1:01:59

to inspire someone as a big

Robert Best 1:02:02

yeah.

Milosz 1:02:02

What would that be?

Robert Best 1:02:04

well, I guess you just mentioned one of them, which was, you know, our, I mean, RNA biology is, is a big challenge, um, understanding of RNAs fold and interact with their partners. Um, another one is, I guess, transcription regulation. That's, you know, obviously, um, something that's very complex and all different factors, um, hard to really address by any one method. I think obviously, obviously very important for many, many things. So, um, there would be an interesting problem to the future and of course, yeah, I mean then there's going in the same direction as transcription regulation, I guess there's the genomic structure, high order structure, and how is that regulated

Milosz 1:02:53

Okay.

Xuhui Huang 1:02:54

So can I add one point, before we wrap up, so then we talk about for student, I. Entering the field now for the students that are close to the graduation and they're thinking about their future career. I think another point I want to add is, uh, there's a lot of the, industry companies working on the ai, AI for bio physics, AI for drug discovery, and a lot of these companies are taking this, uh, data driven approach. I think it's a, it's a good idea for students get exposure to that maybe they're doing an internship in the company and they, they will help them to make informed decision, uh, what they really wanna do in the future after they obtain the PhD. Do they want to work in a company or do they want to move forward in academia? Uh, so that's, that's my experience in recent years. Yeah.

Milosz 1:03:40

Or maybe we should be the the ones starting, the companies

Robert Best 1:03:45

Also, it's not a one way path necessarily. So, uh, people often think of.

Xuhui Huang 1:03:50

But I see the students really are, struggling about, they're thinking about what's the best way to go.'cause they're seeing these, uh, companies in the industry, right? They don't really need to know physics because everything is data driven.

Milosz 1:04:00

That's interesting point.'cause yeah, we try to highlight many times on the podcast that this is a two-way street that people can go learn and go back. Uh, but there's definitely a strong focus on the data science part rather than the biology part or the physics part. Right. So yeah, it's one of the three pillars, Epil r but you can definitely develop there. Okay, wonderful. So Xuhui Huang, Sonya Hanson and Robert Best, thank you for talking to me and for exchanging the ideas and uh, I hope we inspire some young souls. Going into biophysics was a great discussion.

Robert Best 1:04:39

Okay.

Xuhui Huang 1:04:40

Thank you. Thank you.

Robert Best 1:04:41

Thank you.

Sonya Hanson 1:04:41

so much.

Milosz 1:04:42

uh,

Sonya Hanson 1:04:42

was great.

Milosz 1:04:43

a great day.

Sonya Hanson 1:04:44

You too.

Milosz 1:04:44

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