Episode 27 - Kresten Lindorff-Larsen: Refining force fields, the science of intrinsically disordered proteins, and writing better grant proposals

In episode 27, Kresten starts by explaining his path from a wet lab biochemist to a computational biophysicist, a story full of open-ended explorations and helpful mentors. He gives us some background on how both the legacy and latest models developed, highlighting how in each case the driving force were experimental results that either weren't quite matching simulations, or were plenty enough to allow for top-down training. We walk through some of the functions and applications of intrinsically disordered regions, or IDRs in short, and their relevance for medical research. Then towards the end, Kresten shares some tips and observations from his work in grant evaluation, insisting that internal peer review remains the best source of feedback, but in the end it's one's scientific intuition that has to guide us.

Transcript

Milosz: 0:00

Welcome to the Phase Space Invaders podcast. Today you're listening to episode 27, where my guest is Kresten Lindorff-Larson, a professor at the Department of Biology at the Lindstrom Lung Center for Protein Science University of Copenhagen. Many of you will know Kresten from his work on force field development, in particular the very popular Amber 99 ILDN. But his work extends deep and wide from integrating NMR data into protein ensembles, to studying the dynamic folding and unfolding of protein domains to modeling intrinsically disordered proteins or protein regions In particular, Kresten's work in the latter direction has been recently thrust into the spotlight with the recent course Grain Model Calvados, that brings us one step closer to capturing the properties of unstructured regions. And again, that's far from a complete picture of what he's been working on. So we start by revisiting Kresten's path from a wet lab biochemist to a computational biophysicist, a story full of open-ended explorations and helpful mentors. Kresten gives us some background on how both the legacy and latest models developed, highlighting how in each case, the driving force were experimental results that either weren't quite matching simulations or were plenty enough to allow for top-down training. We walked through some of the functions and applications of intrinsically disordered regions or IDR in short, and the relevance for medical research. Then towards the end, Kresten shares some tips and observations from his work in grant evaluation, insisting that internal peer review remains the best source of feedback, but in the end, it's one scientific intuition that has to guide us. I hope you like our conversation today. So, Kresten, Lindorff-Larsen, welcome to the podcast.

Kresten Lindorff-Larsen: 2:04

you very much and thanks for having me.

Milosz: 2:07

So, Kresten, very recently, made a bit of a splash with the exciting research that out on intrinsically sorted regions with your latest Calvados model. But you are one of those people who really leave their mark on our field of computational bio physics in so many places. And, uh. You know all that. Starting as a biochemist, not a physicist, although you've been working very close to physics in many of those projects. So back to your days at the Lab of Michele Vendruscolo, by the way, a former podcast guest, did you find this transition natural or daunting or exciting?

Kresten Lindorff-Larsen: 2:44

Thanks, uhI think all of the above. I was thinking, you know, through this earlier today, and I really think that a lots of the things that we do have happened because of chances. So I, I studied, as you say, biochemistry and I did undergraduate work doing wet lab experiments, purifying proteins either from yeast or bacteria or from barley grains. Uh, and then I went to Cambridge to do my PhD with,, Chris Dobson. And Chris had just moved from Oxford to Cambridge, and so the lab was, pretty non-existing at the time. And so I started doing wet lab experiments there in the beginning, but it was also clear to me that I wasn't really making so much progress, because the lab was, pretty empty. So I started talking to Chris and other people, at the department and ended up, you know, talking quite a bit to Mikala and then did lots of work together with Mikala and Chris during my PhD. So this was really a, bit of a random point. Uh, and then I just had to sort of learn the, the biophysics and the simulation tools along the way. Um, so, very much by chance. Uh, and then I've sort of have been trying to catch up ever since.

Milosz: 4:02

Yeah. Failed experiments account for a lot of great, theoretical computational biophysicists in the end.

Kresten Lindorff-Larsen: 4:09

Yeah.

Milosz: 4:09

Do you have any tips for people who are doing this transition? Like how did you, how did you manage that? Was there a structured strategy for this conversion, or was it just

Kresten Lindorff-Larsen: 4:20

I. mean, I was very lucky to have great people supporting me. So Chris and Michele, Emanuele Paci, all helped me out enormously in the beginning. And then I had already during my undergraduate studies. Studied quite a bit of physical chemistry, uh, and statistical mechanics and some quantum, chemistry also. So I was at least used to some of the more technical things and, and had a reasonable background in some of the theory. And so that made some of the things, easier. but then I think the other thing that I did, and you know, this was not really my own decision, this was again, a little bit random. And was that I ended up focusing on things that sort of bridge between the things that I knew and the directions I wanted to head. So since I had a background in biochemistry, it felt very natural to me sort of to, try to work at the interface between simulations and experiments. And this, of course, was again, something that. Chris and Michele and Emanuele and, and others had pioneered a few years before, and so I ended up sort of in the right place at the right time, and not necessarily with the right background, but trying to, to build the knowledge that I needed to do these things.

Milosz: 5:36

Right. It feels like the interfaces are where of the action is happening. That's, that's

Kresten Lindorff-Larsen: 5:41

Yeah. And, and then the other thing that I think was, again, help, this was sort of a very practical thing, which is I had a scholarship from the Danish government, uh, to go to Cambridge to do my PhD. And so I had quite a bit of freedom in. What I wanted to do, which was great, including that I decided to spend my summers in Copenhagen and so I arranged to have a a desk. Uh, together with one of my former supervisors, professor Flemming Poulsen. Um, and so I sat and worked from their lab, over the summer. And so that lab was doing an, my experiments on amongst other things, unfolded proteins. And so I ended up talking to and, and others around the time who are then doing, you know, my experiments on things that I thought exciting and so, because I could spend some time in Copenhagen, I ended up picking up these, you know, experiments and experimental data and bringing the data, to this joint project that Cambridge. So again, this was completely by chance, but also maybe having the background to see that there was winding possibilities in doing this kind of collaborative work.

Milosz: 6:51

Right. I see. And that was also the beginning of the, golden era of protein folding. Right. So. You ended up at DE Shaw, soon after that, working on the actual folding of those small protein domains or like fast folders, and also contributed to the ILDN, Amber Forest Field, which is still a cornerstone, like probably still my favorite force field to, to use for protein simulations. so how did that unfold at the

Kresten Lindorff-Larsen: 7:19

Yeah, that's a, that's a good point. So I, I was just looking it up and it is, you know, exactly 15 years ago since it was published. Yeah. Again, this is a little bit random. so if we take a little bit of a step back, I worked on protein folding quite a bit during my PhD. I was fortunate to be at a department, you know, with Chris Dobson and Ellen First, Jane Clark, and many other protein folding experts. And so this was always an interest of mine and then I sort of got sidetracked into first studying the Unfolded state, of, of protein, sort of trying to start the starting point as well as the transition state. And this was then work that I took with me when I moved back to Copenhagen after my PhD. so we work quite a bit on studying both unfolded proteins and, and then also intrinsically disordered proteins that we can make return to. And as part of this, we developed a, a method quite early on learning force field parameters from experimental data. So we published a paper, you know, one of the first papers when I started my own group after my PhD on. Using sort of Bayesian approaches to learn force field parameters against the experimental data in sort of a top down approach. And so then when I moved to New York, that was one of the things I was interested in. And I started, a program where we tried sort of to set up these approaches and in the end, some of those things work pretty well. But again, a little bit by chance we ended up being sidetracked into looking at these. side chains and found that there were some of the side chains that looked to be not quite as, you know, as, as good as we had hoped. and while I would've loved to have done this in sort of a top down approach using, you know, my background in experiments, we actually ended up privatizing it by quantum methods. Um, so we have, you know, good colleagues that are experts in quantum chemistry and then using a broad set of experimental data as validation, and so this was all again, you know. Because we knew that working on force fields would be important if we're going to run these long simulations. We started to see some problems with deviations between simulations and experiments, and then use that as the starting point for tackling, these force field problems. So again, you know, and I think this is probably true for most people working on force fields. This comes out not of an intrinsic desire to work on force fields, but because one discovers some problems with a force field somebody is choosing and then you, you have to fix, you know, fix them in one way or another and to do the research you'd like to do.

Milosz: 9:57

Yeah, it's a constant back and forth between applications and methods, right? The best methods come out of applications, and the best applications are what is at the forefront of the current experimental capacities. So yeah, that's a common theme across your career, which is, uh, I think a great template for people who want to have their own. fulfilling scientific careers. And so now we ended up, with, working force field for idr, right? Like, okay, you have this Calvados model, which is already in version three. I, I remember we had a conversation about naming conventions and how you want to version, iterations of the forest field. So let's talk about Calvados V3.

Kresten Lindorff-Larsen: 10:37

Yeah, it's, uh,

Milosz: 10:38

Um, what, what can it do? Like, what, what can it do that we couldn't do

Kresten Lindorff-Larsen: 10:42

yeah, don't know, to be honest, it's, you know, it's again a project that sort of started around 2000 and. Five or six with this work that I mentioned before. So one of the wonders of being at a, at a university with lots of teaching is that you get to work with great under graduate students. And so I had a, a master's student Annes who worked together with me and then a, a colleague at the, at the Niels Bohr Institute of physics. And we then got to start doing this early development of learning force field parameters from experimental data. and back then there was just not enough data really to do this with the breadth that we, we liked. So we did some work with, developed a course grain, force field, but there really was very little of the kinds of biophysical data that we would need. And then I sort of got sidetracked or worked unfolding and, and, and, and other things. And then some years ago, again, I thought, you know, maybe it would be time to, to pick up, uh, this project again. Ramon Crehuet from Barcelona, came on the sabbatical in Copenhagen. There was a, a great again undergraduate student here, Klarsø Schulze who, who worked with Ramon. And we picked this up and then Giulio Tesei came into the group and, and joined the effort. Really just trying to make a basic model that we thought could, be accurate. And I think sort of the driving idea was, and that's been true for many things, is that experiments are in some sense, infinitely accurate. Of course there can be, I. There could be uncertainties in the measurements, but at least they report on nature. And often because you know of,Boltzmann n, factors, you can see deficiencies or small deficiencies in force fields. And so tackling, you know, targeting experimental measurements can really fine tune things quite a bit. So we had the idea we could do a, a relatively accurate coarse-grained model on a restricted set of problems, uh, which initially was just to capture long range interactions in disordered proteins. That was sort of what now we call Calvados one, although we didn't actually call it Calvados at the time. And then we had sort of a fine tuned version that we then ended up calling Calvados two because, you know, rather than do more fine grade naming as you mentioned. Um, and, and, and I think this model is still great for, for studying disorder proteins, but of course, uh, disorder proteins is. An extremely broad term, but it covers many, many different things. And disorder proteins is not just one thing or one biological function. And of course, they generally exert their functions together with folded domains and other macromolecules, RNA, and so. It was pretty clear to us that if we wanted to have a model that we could use more broadly, we needed some way of bringing in the folded domains. And so the Calvados 2 work came out of that desire. we initially thought it would be easy. Then our initial ideas turned out not to work. Uh, but again, JU and worked on this and figured out sort of a relatively simple hack that allowed us to bring in folded domains, at least in some, some primitive way. So these models are not perfect. There is a restricted set of things that they can do, but for those things like characterize global structural properties, they work surprisingly well for a model that has one bead per residue, uh, 20 parameters, plus a simple model of. charge and electrostatic interactions. And so yeah, we are having a lot of fun with this model. It's, it's really enabled us to do many things that we couldn't do before.

Milosz: 14:25

I see, I can imagine. Uh, one of my experiences is that coarse-grained models work so much better with all the force field development tools that are supposed to be exact. Compared to all-atom models, right? Because with a lot models you have so many issues with sampling, with sensitivity and noise, uh, and course grade model, I believe, are just much better behaved in all those regards. So yeah, they're probably the best things to parameterize if you have a bunch of experimental data and, to how hard it would be with an atomistic

Kresten Lindorff-Larsen: 14:56

I mean, so we, we've also done that, right? So some of the work that, in particular Stefano Piana and I, uh, were doing, at DE Shaw research and later picked up by Paul Il and others was of course targeting. Parameters and all that, some force fields against experimental data, either directly or indirectly and with, you know, either, a little bit, trial and error or with systematic approaches. But you're absolutely right. It's, it's harder not just because of convergence, but of course, because when there are more parameters, you need to make more decisions about where to fit. And of course, this is one of the problems of top down permitt is that you, you may be. Fixing the errors with the wrong terms. And of course for cost green, mild, where there's just fewer terms, there is an, in some sense, less risk of messing up because there are not so many terms to fix. So these are per definition, of course, effective parameters that capture many different physical effects. And as long as you know that and you try to restrict the problem set and have enough data. I think that, as you say, these things are, are at least more robust for this kind of top down approaches. And of course now lots of people have combined bottom up and top down approaches, sort of in the martini framework, but also of course now with, you know, with the automated all had some, you know, force field prioritization that that several groups are working on and doing, you know, tremendous progress on.

Milosz: 16:26

True. I talked to Paul well before in the podcast and recently here as well, and he. Saying that, yeah, they have a few things that only optimistic forest fields can do in IDPs, but he also recognizes how much sampling is needed for any kind of meaningful outcome to come out that. because of this coarse grained approach, you could move on to what you called, proteome level research recently. Right. So like to look at the properties of the entire human proteome in a single study, which is impressive. Can you tell us what comes out of such studies? What kind of statistical trends or properties you can

Kresten Lindorff-Larsen: 17:03

Yeah, so we, we were very happy with this Calvados model in the sense that it's relatively accurate for certain things. and, and as you say, it's also, pretty fast. And so we started thinking about. How can we use this speed, uh, to do new kinds of things? And, and we sort of took it in two different approaches. One is a project on IDP design that involved running many simulations, but one after another. And the other one is the project you, you mentioned where we, we studied, all IDRs for the human proteome and VA simulations of. You know, some 30,000, uh, sequences. And so I think this is, now, this was certainly a fun project. It allowed us to bring many different expertises in the group together because it allowed us sort of to bridge the molecular biophysics with, you know, more bioinformatics and, and even disease biology, things that we've also been working on. So in some sense it's sort of enabling us to. To do molecular biophysics, but at proteome scales and bridge it in with, you know, sequence based machine learning, studies of disease variants, studies of functional effects of idr, and, really asking things at a, different scale than what we've been doing previously in our research. And this was a, you know, a, a fun project and we are doing more of these kinds, but I also think that this is something I see. Many other people in the field moving in that direction. still trying to work on individual systems and studying them at depth, but also trying to scale up to look at. You know, whole classes of, proteins like we just saw in a wonderful paper on simulations of, a big chunk of G-protein-coupled receptors, studies of, many kinases or in our case studies of, all disorder, uh, things from human protein. And then of course you can then go back and look at this data set as a whole and try to ask questions about the trends. Across this and what is the relationship between sequence and conformational properties, for example. Then we and others have used these things sort of to train simple or, or for, in other cases, less simple machine learning models to predict these properties directly from sequence.

Milosz: 19:26

All right. That sounds exciting. And also opening the way for new biology, right? I studied, uh, DNA binding domains. and it was always, there was always this feeling that you have a small domain that does something that binds or doesn't bind, but. Then you look at the whole sequence and it's like, it's a monster that has a 50 amino acid domain and then 300 amino acids in this IDR, and that's connected to something else that connects yet to something else. Right. And like I always say this question, okay, I'm studying some tiny, tiny fraction of this whole process or the whole structure that is being. Assembled there, but then, you know, how does it really translate into, into the biology, right? Are we starting to uncover this or to like, for example, in, in the field of, I know transcription factors or RNA binding proteins.

Kresten Lindorff-Larsen: 20:12

think a lot of people are making, great progress on this. we are doing some work in this area and, and lots of other people are trying then to bridge these things from the molecular level to the interactions to studying. I. Things that, uh, that, you know, DNA binding at the, at the chromatin level and and linking that is to sort of more general effects on gene regulations. So, yeah, I mean there's this whole hierarchy of, of interactions and, and, and regulation mechanisms that I think we're trying to, you know, as a field to uncover, but I will also say one of the. Wonderful things of working at a Department of Biology that's extremely broad, is that I get to see, you know, on the firsthand basis that there's always another scale upfront, right? So, we are sometimes, you know, asked, why are you studying these things, you know, in a computer, right? Biology is much more complicated. You should do cell experiments, and then if you do cell experiments, you're asked to do it, you know, in an organoid or in a, in a mouse. So. There's always a scale above that, that then there is, you know, again, there's ecosystems and interactions between ecosystems and so I don't think that there is a single scale that, you know, covers everything, but there's complexity, you know, in, in many layers up, and in many layers down. And as long as you realize that and have some understanding. Of where your research fits in. I think that, you know, studies of individual proteins are still extremely important. Uh, and then you just either need to know enough or, or collaborate with the right people. I. To put that in the context for say, as you say, understand how do these proteins interact with co regulators or, how do the transcription factors bind the DNA and bring in other things or open up chromatin. think that all of these things are necessary and I also see that lots of people are now trying to, you know, really bring these things together in sort of a more holistic description. I think that's just wonderful.

Milosz: 22:11

Yeah, it definitely goes all the way. do you have any interesting examples of function of IDR that goes beyond just, you know, connecting two domains, did you discover or uncover something that's really, really interesting in terms of, you know, why this IDR has to be what it is to perform its biological function Exactly. Rather than just keeping things in proximity.

Kresten Lindorff-Larsen: 22:34

it's a good question. And, and, and, and my initial answer would, be no. not because there's not individual IDRs that we like, but, you know, but I think one of the things we also like is really to look at, things across. Uh, across scales. So we do work on individual proteins, often in collaborations with people that, you know, then do experiments or I'm working with colleagues here, department on, on RNA binding proteins that are involved in epigenetic regulations or in transcription factors. and then I, I'm in the wonderful position of getting to see glimpses of all of this interesting biology. but for myself, I'm really focused on, on these things as a general class of, of problems. And then I also think that while, and maybe this is, to some extent, the IDR Fields, fault, But you know, IDRs very generally do things together with the folded domains, right? It's not like one can always see them in isolation, right? So in our paper we looked at them, you know, very concretely in isolation, and of course this is a major limitation of that work. but yeah, they do things like, as you say, they bring in, uh, folded domains together, right? We worked on multi-domain enzymes where the relative orientation between binding domains and catalytic domains determine activity or as you mentioned, you know, transcription factors. Clearly the DNA, you know, DNA binding domains play roles in binding to the DNA and then the, the activation domains help bind other things. Um, and so I think one has to sort of look at them as a whole. But of course, many of these things don't work without the disorder or the regions. Right. But they're complimentary and do different types of interactions, than their folded domains.

Milosz: 24:18

All right. So when you look at, um, say mutations in idr, what kind of property changes would you be looking

Kresten Lindorff-Larsen: 24:25

Yeah, so this is something we've looked at quite extensively and, Maybe not. So surprisingly, you think of folded domains vs a disorder domains we and other people have found, uh, that, you know, missense variants that cause disease generally, of course enriched in the folded domains. They're much more fragile than the, than the disorder regions. You can have many individual, uh, amino acid changes in disordered regions that have very small functional effects. And then if you look at the disordered regions. We and others have found that they are sort of enriched in the, in the regions that, that probably then make specific interactions with other proteins. In, in short linear motifs, you can sometimes discover these with alpha fold. Alpha fold sees them as being more conserved. so there is about 15% of the disorder regions that may be fold upon binding and, disease variants are enriched in those. And then outside of those there are. Lots of mutations in total number, I mean thousands of them. But I don't think we as a whole, as a community really understand whether there are general mechanisms or whether there are just lots of individual cases. So mutations that affect post translation modifications. I mean, there's the, the, the mutations in the short linear motifs that I just talked about, but if you take those away. There may be mutations that, you know, increase aggregation. There is of course known in, in a bunch of diseases missense variants that caused perturbation, for example, to the formation of, of biomolecular condensates. But again, they are tend to be relatively robust. And so, for example, in one of the many supplementary figures in our paper, we sort of looked at the changes of the biophysical properties of missense variants that cause disease versus those that don't. And there was a statistically significant difference but, you know, not really a difference that explains much of disease variants. Um, and so I think we really need to, have a much better understanding on what are the interactions that these things make, what are the relevant polymer properties, uh, that are important, and then use that to tease away biology. It could also just be that there are few general mechanisms. So I think our work on folded proteins have shown that there are very general mechanisms, right? About 50% of missense variant folded proteins, they just mess up the protein structure and cost degradation. And maybe there are just not so many simple general mechanisms in disordered regions, but I think time will tell.

Milosz: 27:08

Right. Seems like the next, maybe not so low, but hanging fruit,

Kresten Lindorff-Larsen: 27:12

Yeah, I think,

Milosz: 27:13

to grab

Kresten Lindorff-Larsen: 27:13

absolutely. No, I think, but I think that there's lots of people working, on this. And of course we are now starting to accumulate, you know, more and more data and corresponding more and more, simulation and, and other computational methods that can start to address these things. So both predict which variants cause disease, but also trying to peel away the molecular mechanisms of. how they cause disease. And I think again, you know, these kind of models like Calvados and others, I think will hopefully play a, a useful role. I think in the IDR field we are still, a little bit, uh, hampered by the fact that we don't have the same amount of systematic data, uh, and tools that work so well for folded proteins. And so we need these kind of physics restrained models to interpret the experimental data. This can either be simulation models or theoretical models or, or, or combination of these for the machine learning models that then allow us to look at, you know, look at these things at scale, but learning from, limited or imperfect data in a way that, you know, maybe one can get away with not doing, if you're looking at the folded domains. But yeah, I'm, I'm very hopeful and I think it's also fun and, good that many people are working on this.

Milosz: 28:30

Right. So back to incremental improvements based on whatever is available. that's a common theme. yeah. Also, quite notably, you can now design IDRs, right? So you can use, this model also to generate IDR with specific properties. Why would you want to do that? Like what, what is the selling point, for bespoke idr, for example, or IDPs.

Kresten Lindorff-Larsen: 28:50

Yeah, I think that there are a couple of of reasons for doing that. One of them was, you know, just the fundamental curiosity question, can we do it? And then building on that saying, well, if a third of the amino acids in the human proteome are disordered and there's lots of evolutionary and genetic and biochemical evidence that IDR are used extensively by biology, they must be able to confer functions that are maybe not so easily performed with, with folded domains. And so say if biology uses it, wouldn't it be great if we could also, design these things sort of at, will. And of course, people have, designed IDRs before using, uh, sort of rule-based design methods and. They work very well. And what we sort of tried to do was say, well, maybe we can take sort of a more, automated, computational, method where we sort of let the, where we, the simulation tool design the sequences either directly by these brute force simulations, as I mentioned, or, you know, using machine learning models as intermediate models. And I think it's an open question on, you know, what exactly we'll, do with them. I mean, we. We've not really done anything. practical with these things yet beyond test our understanding of sequence, example, function, relationships. But again, you know, if you want to design enzymes that target specific surfaces with, you know, many of them, they have, you know, long disordered regions that tether together substrate binding domains and catalytic domains. We have lots of receptors that have disordered regions. You know, maybe if you want to design orthogonal you know, signaling molecules, having some design protocol would, would be interesting. So I think it's really, you know, an open field that is starting to, take off now with, multiple groups kicking this back over again, both for sort of curiosity driven research and this tool and starting biology and maybe for sort of technological, you know, applications of using them to, to improve enzymes or, or formulation or, other things. but I think, you know, we, in this case we sort of started the project, see both, can we do it? And secondly we thought, well, if we can design proteins that do not exist in nature, and then we can go make them, we can then compare them to, the model, the simulation model, and then use that as a relative stringent test on whether the model extrapolates well to completely different sequence system from what we we on. And I think. It was nice to see that. Yes, indeed. It worked very well. Also for things that look nothing like, uh, like the sequences we have in nature,

Milosz: 31:34

Right. That's a great point. we expect to see attempts at, uh, IDP therapeutics in, five years from now?

Kresten Lindorff-Larsen: 31:42

I mean, to some extent the success of peptide hormones as drugs is an example of this. Right. Many of these are, very flexible. Uh, I come from a country that, uh, produces lot of GLP one receptor agonists, and so, uh, these are certainly disordered, molecules. Uh, and of course lots of, of peptide hormones are like that. I don't know whether we will see IDP drugs in general. You might see drugs targeting, disordered regions or interfaces between disordered regions and, folded domains. but I don't think we had a situation where we necessarily want to, uh, to develop IDP drugs. But of course, someone will, then tell me that I'm wrong and, and, and go do this. And that would of course, be great. No, I think that, uh, you know, we now can design. Folded proteins that are basically whatever structure we'd like, and there's much progress in also making them bind to other things and maybe even work as catalysts. And so that's of course solving a big part of what proteins do in in biology. But then of course, proteins are dynamic and so people are working on putting in dynamics into folded domain so you can switch them on and off. And I think disorder proteins is just one of the tools or disorder regions is one of the tools that biology uses to, regulate other things. And so if we can, you know, design those as well with whatever properties we want, then we can, potentially expand the scope of protein design to sign to all the same kinds of things that the disorder proteins do in cells, whether that be as enzymes as I mentioned all in synthetic biology, as you know, transcription factors or.

Milosz: 33:29

Right. It, it definitely goes both ways. As you mentioned, that people working on. drugs that bind to IDPs or IDR as well. And I don't know how promising I feel that is, but it is a field now,

Kresten Lindorff-Larsen: 33:41

I, uh, it's a, you know, of course I have a conflict there because I, I work with people that do this. But I will say that I think there's so many important proteins with disorder regions that if we could do it This would open up lots and lots of possibilities. But as you say, and as many other people have said before, right, it is of course an incredibly difficult endeavor. We haven't really explored it that much and very systematically yet. So I think it's still pretty open how far this can be pushed and what the, the tools and the ideas are. And I, you know, I think it's great that someone is doing this. I think these are the kind of problems that. We will only really find out by trying. I think it's very difficult to predict whether this will be successful, or not in general, and of course in individual cases it can be very successful.

Milosz: 34:32

Right. It feels like, uh, as you say, it's hard for our community to tackle it. Maybe, it's easier to do in those broad screenings, that are done experimentally. I mean, you have a lot of biochemists and experimentalists around you, right? So maybe, it's a question to your collaborators rather than to. Computational biophysicists.

Kresten Lindorff-Larsen: 34:50

Yeah. Or, or I think as with so many other things, as sort of finding the right interface between computational methods, simulation methods, machine learning methods, and, and the right kind of experiments that have the accuracy you need and the scalability that you need. And again, you can do lots of things with phenotypic screens. but then it's very difficult to figure out exactly what you're hitting and the mechanisms. So I think all of these things, they really have to, come together to address these, these questions. And that's, you know, one of the wonderful things about having the jobs that we have, I. Just.

Milosz: 35:25

right, I can clearly see the role of computational physics in say, lead optimization, but not maybe lead design or like coming up with the initial scaffold for something that binds to something un structured, right. But you can probably guess what you can add to a molecule to make it, uh, stick more to some,, kind of stable mode of

Kresten Lindorff-Larsen: 35:44

Yeah, I don't know, but, you know, I'm very happy that someone is, is trying to, tackle this and build the data and the understanding and the tools, uh, to do this and, and, and clearly our coarse-grained model working at, you know, the resolution of one bead per residue. Doesn't have in itself the chemical accuracy that's needed to study, you know, the effects of individual hydrogen bonds. But maybe together with, uh, other tools, these, things can, play a role. And then I think of course, that the work that we and many other people are doing on trying to, you know, look at biological problems, uh, at a larger scale, hopefully can also help with the stance of general rules of where to target things or, what are the kinds of, sequence motifs, specificities one might, necessarily, uh, you know, where could potentially address.

Milosz: 36:36

Yeah. right tool for the right question. you alluded to that before, that when you do grant evaluations very often, this is a. Huge consideration that people often skip, right? Like they try to approach a big question and then they end up, tackling, completely unrelated, uh, sub field or sub area of that question. what are your guidelines or suggestions or tips for people who struggle with or generally write grant applications, to make them Both having a higher chance of success and maybe making it better at addressing actual

Kresten Lindorff-Larsen: 37:12

Yeah, I, it's a difficult problem and, question, right? I, think lots of people will, will tell you, including me, that you should try to put yourself in the, you know, in the shoes of the person who will read it. which of course already brings up the first problem, which is that you don't know who will read it. And secondly, we all individuals with our own. pet peeves and niches and expertises, right? And so there is some randomness in this process simply because the people who will read the grants, even if it's multiple people, will be individuals with, with expertises and, explicitly or implicitly things we, we like and dislike. but that said, yeah, I think it's important. to try to, to write things so that you explain why what you're doing is important. I think that's sort of, you know, in some sense obvious. But then I think what I sometimes see lacking is then the connection between the actual proposed research and the bigger research questions that you set up. Right? So in some sense, you know, you write your, your introduction and your overall. Research area names, and then there will be some middle part that explains how you will do this. And then as an evaluator, I would like to see that if everything makes successful, at least, then that should provide some, some insights or progress towards the big question that you set up in the beginning. and sometimes I find that this is, can be missing that people, they, talk about some big. Overall problem in biology or physics or whatever, the topic is. And then they go right about what they're going to do and then they don't really help, you know, make the connection to how this will move their field forward. And of course, sometimes I think it's because people don't think it through. I think more commonly it's just that they think this is obvious, but maybe it isn't obvious to a reviewer who has a slightly different background or who doesn't know the field. Um, so I think this is one of the more common things that I would, I would encourage people to think a little bit more about, sort of, you know, really connect the whole arch from the idea to, you know, to what you actually do, to where this will move the, the field, uh, forward. And then in general, you know, make it an easy and reasonable thing for their reviewers, to read. But I will say that I, it's a lot of time it takes, but I really enjoy reading grant reviews. I learn a lot both about science and, and how people think. and also I think, you know, learning how people are trying to connect, uh, ideas. And it's also good to see that there are many ways of writing good grants. There's just not, you know, there's not a single way that this, you know, must be done. And I think this is always useful to sort of see. See different ways of how things can be done because of course, I, I tend to write things my own way and, and that's not necessarily, uh, the right way. So that, sort of some high level views. And then, you know, even if one is, successful in the sense that you are, you know, running with the average success rate, it still means that most grants that you also bid will be rejected. And so the other recommendation I have, you know, involved is just to submit many grants, which, can feel frustrating. Um, but with the way that things operate, that's, probably the way that you, we have to go for a while.

Milosz: 40:24

Yeah, this is how you win the lottery. You

Kresten Lindorff-Larsen: 40:25

Exactly.

Milosz: 40:29

It's a fair point. But also, I think, well you have this cross disciplinary, experience and I think a lot of people in the field l perhaps me included. I mean, You know, it's pretty intimidating to go from a computational model to like an actual, biological outcome, right. For most of people who are on the theoretical side. So this is a question of how we build this kind of market, place of ideas where people can give each other feedback and like improve or add to. Should grant, for example, have you think, should they have rounds where you would improve your proposals with each feedback? cause sometimes it feels like people just submit the same application and, you know, and it's either successful or it just goes down the drain, right? Like how much you could maybe improve if grant applications were more like papers that you submit and you revise and, uh, maybe something interesting comes out of the

Kresten Lindorff-Larsen: 41:26

Yeah, I think that certainly you can get feedback that can improve things. Of course. These kind of things exist to some extent in different funding bodies, uh, in the sense that you do get feedback and sometimes, you know, it's even useful, maybe not always so useful. I think then it runs into the. problem that of course next time you, you submit, it might be a different group of people that review it or you might be competing with a different set of alternative, proposals. And so something, you know, if you're told, well, this is great, but if you cut away this part and that this other party would be even better. Then if you do that, you might end up with a different reviewer that, or set of reviewers that say, well, I don't really like this. Or they will say, this is great, but there were others that were even better. And so I'm not sure that there is a simple way of doing that. So I think at least this is what I recommend is that. Yes, it's great to get feedback, but the best way of getting feedback on grants is to show your grant proposal to your colleagues. and maybe not just the colleagues next door, although that that can be great, but maybe also someone who is I. You know, a little bit further away from your scientific discipline. Now, in my case, you know, the colleagues next door are sort of, you know, a little bit further away, but I think it's both good to get some detailed feedback, but also from someone you know, who's a little bit further away. Because, you know, in most cases the reviewers will often be people who are not experts on the things that you like And so I think it's always great to, get feedback, on your proposals and ideally from as many. people as possible. We established something here in our department where we were a group of people that offered to provide feedback to, grants. of course, people end up doing things last minute, which of course makes this hard sometimes. me included. but I think this is something I can really recommend, to get feedback and, and to listen to it. But also at the end, trust your own instincts, because of course, you are also yourself the expert on your projects

Milosz: 43:34

right, so peer review stands the test of time

Kresten Lindorff-Larsen: 43:37

I mean, it's it a good mechanism and to be honest, sometimes it just makes you rethink. things like to say, you know, you, you, you thought you wrote something perfectly clear, and then someone, who should know this reads and then they don't understand what you're writing. And it's clearly, because, you know, you made assumptions on behalf of the reader that were not correct. And I think it's just useful to get that upfront. I will say though, that, you know, you mentioned these things of interdisciplinary research projects, and it is hard, and I've seen it upfront, both when I've sat on the ERC panel and also in my current panel, is that these projects can sometimes have a hard, time. I mean, they can also fly through because many people are excited about it. but I think, you know, if you, if you don't really fit the target audience. Many people that may think that this sounds interesting, but maybe don't, are not really enthused about the, the specific details of that project. we had one project that, you know, was funded finally. but that was evaluated as, you know, as an ESC project, both in the physical sciences and the biological sciences. And I think that, I mean, they've got really nice reviews and very thorough. I was, no, I have no complaints about the reviews. they were, you know, really good. But I also got the impression that the physicist maybe thought it was very biological and the biologists thought it was very physicsy and weren't as enthused about it as I were. And maybe it wasn't written, you know, as well. But I, I, I've seen this both, you know, as an applicant, but also as an evaluator that sometimes these things can have a hard time.

Milosz: 45:18

Yeah, I can imagine. I mean, these are probably the best applications most of the time that are sitting in the middle. Right. But then who's exactly, who's, um, tasked with evaluating it? That's, that seems like a hard problem. I don't know if we have. Good

Kresten Lindorff-Larsen: 45:33

No. I mean, I think, you know, of course in all these, we try to evaluate broadly and we will send out things to. Other reviewers if we don't have the expertise. And I'm also not sure that interdisciplinary research projects are, inherently better. There are certainly lots of field specific work that is extremely high quality and very important. maybe if I, you know, can boil it down. I think that one rule of getting things funded is that at least one person on the panel needs to love your grant. Ideally zero people on the panel, hate it. and I think, you know, so this thing about having someone on the panel love your grant can be a little bit more difficult if it's sort of. Falls in between, disciplines, right? If, if you know, then say, well, this is interesting. I'm also interested in these things, but if you, if there's no sort of expert who can really see the value of these things, I think these things can be a little bit, more difficult. I know this is obviously not a, a hard, fast rule observation, but I, see that that's paper where these kind of projects, fail. I will not complain. We've been very fortunate in getting external funding for, for our projects. Also the ones that bridge across disciplines. Um, so, you know, it's, definitely possible. in particular

Milosz: 46:52

Yeah, it can also be, it can also be in the writing, right? This is something that's hard to quantify how good of a writer someone

Kresten Lindorff-Larsen: 46:58

Yeah, I mean, I, I,

Milosz: 47:00

it's hard to give an advice. just

Kresten Lindorff-Larsen: 47:01

uh, I mean, I've reviewed grants on. Molecular biophysics where all the words and all the abbreviations make complete sense to me. And then there's another reviewer that's sitting say, this is all, uh, nonsense. To me, this person is not defining any of the terms why they're writing this. Right? So it's very much about, you know, the background of these things. That's, that's, unfortunately just the way it is. And, and you have to think about that, of course, when you write the proposals. But at the end it is to some extent a lottery.

Milosz: 47:32

Right. And I think there is research that says that actually, um, research that is more sitting across, at least to disciplines, ends up being more groundbreaking or transformative. But, well, as you say, it's not to complain about research that is within the well established

Kresten Lindorff-Larsen: 47:50

and I also, I.

Milosz: 47:52

Brings great results.

Kresten Lindorff-Larsen: 47:54

it makes sense, but of course it still requires that we need some kind of metric for judging what good research is. And sometimes you can do this, right? But, but often these kind of large scale analysis, I think they tend to boil down things to simple metrics of how, you know, how were the papers, how many papers came out, and how were they cited? And you know, that that is certainly a metric and it can be useful. But I think it's very difficult to address these things at a scale where you get statistics without relying on these kind of simplifying, uh, metrics. So I think the only real solution is to keep lots of diversity in funding and funding structures. ideally, of course with as much, you know, funding as is needed to do the right kind of research. but then also just to make sure that we can fund. Big projects and small projects, interdisciplinary research, focused research, research for junior scientists, but also for mid-career scientists who, you know, sometimes struggle enormously. I think, you know, we just need this kind of diversity to keep, you know, people, you know, ability to do their jobs and use their creativity and, and good ideas for whatever it is that they're working on.

Milosz: 49:13

Absolutely no easy solutions, but that's why it's worth thinking about. Okay. Kresten Lindorff-Larsen, thank you for coming on Phase Space Invaders and sharing your expertise and your insight.

Kresten Lindorff-Larsen: 49:24

you very much for having me, and thanks a lot for running this podcast. I, I'm enjoying it a lot, uh, listening to it, so thank you very much.

Milosz: 49:32

Thanks so much. Have a

Kresten Lindorff-Larsen: 49:33

Thanks.

Milosz: 49:35

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