Phase Space Invaders (ψ)

Episode 31 - Ezgi Karaca: Protein docking post-AlphaFold, legacy integrative modeling, and the importance of training events

• Miłosz Wieczór

Share episode
Milosz:

Welcome to Phase Space Invaders. In episode 31, my guest is Ezgi Karaca, who is an assistant professor at the Ismir Biomedicine and Genome Center and Dokuz Eylül University. Already from her early research years in the lab of Alexander Bonvin, Ezgi was deeply involved with the prediction and integrative modeling of protein-protein interfaces. Aside from contributing to the development and popularization of haddock, she was involved in several editions of the CASP competition as an evaluator and as a consequence, her interests have been recently shifting to a critical evaluation of the field of macromolecular docking in the post AlphaFold era, a question that affects us all in one way or another if you're doing computational biology. So our conversation starts at the interface of old and new wondering where do legacy docking methods fit in the toolbox of today's computational biophysicists? We discuss what makes a particular software relevant for the community, both from the perspective of users and developers. Ezgi shares. her observations on the impact AlphaFold and related prediction methods have had in these famous cast competitions and what it tells about their inner, their workings and shortcomings. Finally we shift to talk about the opportunities and challenges of running an impactful scientific lab from Turkey and the importance of training events in a time when paradigms are rapidly shifting. I hope you enjoy listening to our conversation. Okay, Ezgi Karaja, welcome to the podcast.

Ezgi:

Thank you.

Milosz:

So Ezgi, I understand you were there during the many iterations of development of Haddock. both observing and contributing to progress in protein, protein docking and integrative structural biology. And I think so many of our listeners are routine users of these methods and these tools, and we love to hear an insider's perspective on what has changed in the last several years. You know, have AI tools completely taken over. What are the use cases for legacy algorithms and, uh, where do we still struggle to make progress in that?

Ezgi:

okay. So first of all, thanks a lot for your invitation. Uh, this has been a podcast that I have been following. and it's actually a, a very nice feeling for me to be one of the, uh, people who is being interviewed, let's say.

Milosz:

That's lovely to hear.

Ezgi:

yeah, coming back to your question, I think. question has, uh, multiple layers. I'll try to answer them, uh, layer by layer. starting with, Uh, so my relationship with Haddock started, um, when I started my PhD in. 2008 with Alexander Bonvin. So, uh, the main developer and the, owner of, Haddock software, in Utrecht in the Netherlands. so at the time, Haddock was, uh, rather fresh. So its main paper was published a couple of years back, and the, protocol, the idea was there, uh, it was picking up, but it was obvious, not as big as it was today. And it was a time when, uh, web servers, were becoming a thing, a thing in a sense. Like it was a time when people realized like, you should provide your tool as a service. And then it'll be used many users around the road. And, had a web server as well. So, uh, in that context, when I started, the idea, the core idea was already there. the idea of how to develop Haddock further was already there as well. And the bottlenecks, uh, that were there at the time, in the case of big molecule docking, let's say, so not protein small molecules, but protein. Protein or protein, RNA, protein DNA, the first one was, the sampling problem. first, at first, you didn't know whether you sampled enough or not. The second problem was even if you knew how to sample things, you didn't know how to pick the right solution among all the solutions generated. And the third issue was, uh, like since we are dealing with large molecules, how are we gonna deal with confirmational changes, especially the large ones that occur upon binding. So it's, uh, obviously interesting, to list these also today because even at the, um, age of AI or at the age of AlphaFold, uh, we are struggling, uh, with these three, issues in terms of, protein interaction modeling. And I guess we'll come to this throughout the time of this podcast. Uh, but concentrating on this within the framework of Haddock, and which was actually what my PhD was about, uh, was a very smart move on my advisor's end, because, Haddock actually works, differently than the other alternative docking Programs. It, accepts input, external inputs, uh, from the user, uh, which could be, the binding interface, which could be orientation of the monomers, which could be something else, et cetera. And with this actually, if you have such data, you're able to deal with the sampling problem better because then you restrict your confirmational space to a more, real parts. and scoring bits also can be, eased a bit if you add external experimental data to score your complexes next to standard scoring functions. Uh, but regarding the third one, dealing with confirmational changes this. this. uh, this is, and still something that needs to be dealt with externally. Like if you expect a domain motional binding, then you have to generate your ensembles or, dissect the domains before the, uh, docking protocol, and then, feed the ensemble or different domains into your docking protocol. So, uh, this was, what my PhD was about, generally, how to integrate data to address these problems. but then obviously, not only with this, but with multiple other protocols, with multiple other approaches, Haddock became something that was picked up by the community very widely. And, um, I think the main, one of the main reasons for this is that, besides having a web interface, There has been also already quite, uh, some time, workshops organized around teaching Haddock online and in person. I think this is very, very important in computational structural biology to make your software used widely. And second, it has been picked up so nicely because. It is really being shaped by the response of the users. okay. It's an academic world. you can't implement something immediately like people could do in companies. Maybe as user feedback, it takes time. But what I've observed throughout the years, even I, uh, stopped developing things for headache, what I have observed was like sooner or later the main complaints or the main wishes of the users. Are implemented or will be implemented in Haddock. I think these two aspects, uh, made the program, and its user friendliness, obviously, very widely and nicely picked up by the community. So, uh, this, as I said, like your question, content, uh, multiple layers. So one layer is like, okay, we should continue. I think to address one layer, um, I wanted to say like. If we want to have, um, a software to be continuously used by the community a, a biomolecular modeling software, you have to continue growing it. You have to continue updating it and make it user friendly and find a way which is costly obviously, but find a way to advertise it and make it under understandable for the whole community through in-person interactions Yeah.

Milosz:

can, I think it's also a good call for everyone who's using those tools to, you know, give your feedback. Don't just scream at the computer. If you

Ezgi:

Oh,

Milosz:

with something, look who's developing the tool and, and, get in touch. I myself, as. As a software developer, really love it when someone writes to me with a bug report or something that you know they're trying to do and it's not working.'cause it's like I cannot test everything, but you're raising a very good point there.

Ezgi:

Or related programs are dead right now, uh, considering the state of the AlphaFold, just putting it bluntly out. Yeah. they're still needed, but less than earlier, obviously. but they're still needed because in the context of protein interaction modeling with any big biomolecule protein or RNA, et cetera. alpha fold is not able to address all types of interactions at the moment. And this is especially, uh, the case, uh, where the co evolutionary signal is not strong. Uh, like in the case of antibody antigen interaction modeling or viral like host pathogen interaction modeling, viral protein modeling, interaction modeling, et cetera. So I think one can confidently say that in the case of monomer structure prediction, this is the best that it can get for a long time, I guess. for protein interaction modeling, there's definitely, uh, there are gaps still. the gaps or the challenges are the exact challenges that I mentioned, but the severeness of the challenges are different now. That's the main difference.

Milosz:

Right, so those tools still haven't learned the physics as this. This forever debate, I think now in the community, whether they're just memorizing patterns or memorizing entire complexes, right? Or they are learning something fundamental. And I don't know where the, current consensus lies, but I feel it's like the memorizing patterns. And if the pattern, as you say, it's not there for an antibody in the,

Ezgi:

Yeah.

Milosz:

set, is not gonna be there in the prediction.

Ezgi:

Yeah, yeah, this, but I think also what AlphaFold does as an extra is like in analogy with Haddock, it also takes experimental data as input. So if you don't feed any multiple sequence alignment to AlphaFold, We can leave out the template because we don't always, have, the template information. But if we don't fit multiple sequence alignments, uh, into, uh, alpha fold, it's not gonna work. Nice. Even, uh, for monomers. so I think this is a bit in analogy, tad in a sense that if you have good input data in this instance, uh, if you have a multiple sequence alignment for a dimer, let's say. signal wise, it's somehow obvious for the machine of, course not, for us, but for the AI machine. It's somehow obvious that this certain residues should be at the interface and this certain residues should be, defining the structure of the monomer. Then it gives a very good prediction and it gives it very fast, and it's very confident about it. this instance, okay, it doesn't know the physics, that's for sure. But it knows, and I think this has been the great part of this tool. It knows how to interpret the sequences, which is in, in the end also an experimental data. I mean, it's an outcome of genomics efforts. So if, for example, don't know, one could use a denser sequence database, maybe they could be, um, more lucky in some certain cases. But for antibody antigen interactions, for example. I think for AlphaFold or for other docking tools the main reason for their struggle is that also experimentally, we don't know how an antibody really recognizing the target. So the physics are also not there for this, for these interactions. this is a really, uh, an ongoing research effort. There are many labs who are only concentrated on understanding antibody antigen interactions, but with what we have attained, I don't know how successful they can get in the, in the short term.

Milosz:

Right. And there's also the question of actual experimental restraints, right? So Haddock can incorporate things like restraints from cross-linking and so on. There are, I know that there are some AI. Structured prediction tools now that start to incorporate those that you can say, okay, this residue should be within this and this much distance from another residue. But uh, this is not something that's standard, right? This is probably where legacy tools still have an advantage if you have a lot of experimental re.

Ezgi:

Exactly. Exactly, exactly. And maybe some, sometimes even not a lot, some. Maybe in some certain instances you have two instances and some other interaction pattern that you think that should be there, and then you can get a, a good outcome. And it's indeed the case. For example, the recent efforts. in tutorials, uh, from Haddock's end has been, mostly concentrated on antibody antigen modeling. But this also now is a mixture of AI and, Haddock tools in a sense that, for example, you can generate different ensembles of the loops CDR loops by using an AI method. instead of using md, you can do this fastly. Okay, maybe it's not proper, but still then you can use this as an ensemble input, things like this. So I guess, for these difficult systems, it's not, either one of them, but probably to make use of both, uh, in the best way you can.

Milosz:

Right. another thing that's. Also part of those, prediction methods on the AI side are language models, right? So I assume that sometimes you can work with improving the language model that's more informative about possible interactions rather than the MSA. But have you looked into that in any, systematic way?

Ezgi:

Not systematically, but. Just with my current knowledge, I feel like, without having anything extra, like MSA, I don't think that you can really, This to a solvable problem because we don't have enough data for this. So if you think about how AlphaFold trained, okay, It took all the structures in the PDB, but how many structures do we have in the PDB? I think one of the main reasons, why, um, AlphaFold was successful was, that. It is combining an input and output relationship in the end. So that's why I, I just cannot foresee, but it could be also my, um, missing knowledge. But I, I just don't see how this can be achieved with the level of data we have.

Milosz:

I see, because, yeah. Uh, anyway, we get a lot of data from bioinformatics, so all the metagenomics is providing. Tons of implicit data on, on structural relationships. Right. But,

Ezgi:

Yes,

Milosz:

way it is being fed into the, into the models is going to matter a lot.

Ezgi:

exactly.

Milosz:

Yeah. I was wondering if, with the methods that we already have, can we start moving towards something like interactome prediction? do you think we are at the stage where this will become a thing where like we can take an entire. Maybe a prokaryotic organism and figure out, interactions between all the proteins in principle. Of course, not with a lot of detail because some of them are mediated by things other than proteins. But, um,

Ezgi:

Mm-hmm.

Milosz:

you think we're moving in that direction?

Ezgi:

there has been already, um, efforts this direction. I think the first attempt that the fields, uh, tried to do was during my PhD. So, really tried to do this by using docking methods, but more like rigid body, fast, rigid body docking methods. Even at the time that was, that really started a huge, uh, excitement in the field though if you don't know how, if you, if you're not sure about your scoring function, obviously this will generate data and be as much, but that was the best thing that they could do. And it was already quite pioneering, I think. should be around 2010 and 2011. and nowadays, there are already, Two or three papers out that did this, um, maybe even a bit more for a, for a small organism. Also for one I saw recently that I haven't read yet. Uh, one was done for human, interactions. but there, one has to see how, uh, people filtered out things. So I think sampling. Like, you can sample as much as you want with a high-end supercomputer right now. You can try all, all to all combinations though, okay? You have a bottleneck of not knowing the telemetry there, but somehow if you also have, some partial stock information for some of the, some of the systems as sampling is definitely achievable. It's not gonna be a bottleneck. but again, there you have a problem, like, in the case of interaction modeling, we have a feeling for AlphaFolds, how AlphaFold scoring behaves, we are still not sure about how it really behaves, so we don't have a cutoff as in the case of monomer modeling. For example, if you have a value above 0.8, then this is definitely an interaction. An IP TM of 0.8, for example. I mean, probably yes. Maybe not. So this is also an ongoing discussion in the community, but, technically this is totally achievable.

Milosz:

Try because in the end we would like to also predict affinities, right? And, uh, going from those very, very implicit scores. I mean, back in the day, I remember docking software used to have a more interpretable. Scoring that somehow tries to approximate Delta G in one way or another. But with AI tools, we essentially just have this quality prediction. And do we know anything about how good it is at, at trying to recover relative affinities? At least maybe absolute affinities is a lot to ask about. Uh.

Ezgi:

No, I, I don't think so. I don't think that it has this power, but on the other end, it works much better than the docking scoring functions. That, that's for sure.

Milosz:

Right. so can we see the impact of, for example, mutations, um,

Ezgi:

No, no,

Milosz:

on.

Ezgi:

AlphaFold is not meant for this anyway. we can't do this with AlphaFold. But if, so for example, uh, the BioEmu, tool, they aimed so they, um, did not retrain AlphaFold, but. They fine tune it with, uh, MD data and with mutation, with, uh, data that contains the impact of mutation. Let's say they claim that with this protocol they can, predict the impact of mutation, but with AlphaFold because it was not designed or trained for this, you can't do this. You can't predict affinity, you can't predict the impact of mutation across protein interfaces. You can't do these things.

Milosz:

Right. I sometimes it's like.

Ezgi:

claim, to, to give them the proper credit. They never claim to do so, but people just want to get most out of this tool. Right. So that's why, uh, people have been

Milosz:

That's why I'm asking

Ezgi:

Yeah.

Milosz:

Exactly. Because then somethings come out. Somethings come out, but then people come back and criticize that, yeah, this is not physics. But, uh, yeah. So things are actually correlating in interesting ways that, should not be there. Like the only questions that you never have any guarantee. but maybe you have a hint. I wonder how much we can formalize this at this stage.

Ezgi:

that

Milosz:

Probably not much. I see.

Ezgi:

to, you should use AlphaFold. Primarily for a static structure prediction. So this is its primary aim and this is what has been, what it has been developed for in the end. but by using some tricks, you can try to sample different confirmations and this for monomers, not not for interactions at the moment. Or, um, you can try to score your designed monomers. Or interfaces. This was tried, uh, by Sge Nikko in a couple of papers. so this has been shown to work, but beyond this, I don't that you should do unless somebody else will prove it. Otherwise,

Milosz:

Of course,

Ezgi:

you

Milosz:

so we.

Ezgi:

you should not use it for this, for, for the purposes that you would like to learn, actually.

Milosz:

Yeah, so we can introduce variability with one set of methods and then score with another set of methods, and probably that's the most physics, relevant or

Ezgi:

Yeah.

Milosz:

justified way of doing that. Okay.

Ezgi:

A mixture of methods is still a lot. Yeah. Yeah. I mean it, it's just the structured prediction problem is much ease right now, which actually gives us More opportunity, maybe spend more time on some of the questions that we did not have. earlier, like, uh, now you can predict the structure or interaction of x and y, uh, very fast and hopefully confidently. Okay? Then can create your own data set on something very, in a very fast manner, which is great. Then you can also play with any kind of physics based tool, uh, with these predictions. So I guess one should, see the, um, advancement in this way, not a sort of a solution to all problems that we have in the community.

Milosz:

we can start moving towards real biology and real

Ezgi:

Mm-hmm.

Milosz:

hypothesis generation, right?'cause that was always something we could not do as, as computational people. We could just verify someone else's ideas

Ezgi:

That's so true.

Milosz:

to some extent. Yeah.

Ezgi:

Yeah.

Milosz:

a. and exciting time to, to work on this. Yeah. For example, what would you say about last CASP competition? What were the most interesting takeaways there? Because you already mentioned antibody, predictions related to antibodies in antibodies. I, I know that these are hearts, and this was also highlighted there, but, uh, was there any other interesting point that you would like to highlight from that?

Ezgi:

so I was, uh, assessor for the protein complex category for two rounds in CASP. Uh, this was in 2022. Just when AlphaFold emerged, for the TER structure prediction category and there was another round in 2024, but I wasn't involved in this though obviously I closed the, followed, uh, what happens, in the competition uh, it seems like even though in between AlphaFold three was released. the challenges in the case of protein interaction modeling hasn't changed since the previous cas round, since Cas 15. So this, round was the last round was 16. and this also we see in, um, AlphaFold three's paper. So in the, standard protein complex modeling. doesn't seem to be a much difference compared to AlphaFold two, but in AlphaFold three paper, uh, they said like if you run AlphaFold3 one thousand times then you'll get a performance boost. but then in the paper they did not compare what happens if you run alpha of all 2.3 ten one thousand times as well. I think that would've been a more fair comparison, but in any case, running AlphaFold three is so fast, even though you get the exact same outcome. I think running three is way more advantages compared to 2.3. But in complexes, the bottleneck is, multiple sequence alignments, uh, and it hasn't changed, even so since. We experienced the pre alpha fold and post alpha fold parts in person. one can confidently say that the performance compared to the pre alpha fold time in the case of protein complex modeling is, uh, the performance boost is amazing. Like it's, it's really like, times, four times more community performance compared to the previous time. but again, as I was mentioning, the field still has some challenges and I, I would like to maybe, um, mention a nice anecdote that, I had with, John Malt. So ca they have been, into the complex, business since the last, uh, 10, 12 years. So, um, earlier this was picked up by Capri and other competition. They still go on, but Cas was there mostly, uh, focused on tertiary structure prediction. Uh, so then when I was doing the assessment and then afterwards, especially, uh, after AlphaFold, uh, came out and people started to use AlphaFold for Protein complex prediction. I was trying to explain them the change and how, impactful that would be and how, uh, amazing, uh, the performance boosters, et cetera. But even with them, it was at first. Difficult to understand because, you know, protein interaction modeling is something that you can't really, boil down to a single sentence or single hypothesis that would say like, okay, Levita paradox X is paradox, or Y paradox. Because the landscape of binding is very different with different biomolecules, and it's really hard to put down to a single hypothesis. It's maybe even impossible because it's very context dependent, very biomolecule dependent, et cetera. And once when we were, when I was trying to explain them like, how good things are, even though for some things. For some targets, AlphaFold, missed the outcome. And finally they got me, and I remember John Mo telling me like, the problem of the complex, modeling, communities is that not problem, but the fact that the problem is not so famous compared to tertiary structure prediction is that. don't have something like Levinthal's paradox or Anfinsen's dogma, you know? And that really stayed with me. He is right. But on the other hand, this doesn't make the, uh, prediction challenge less important compared to tertiary structure

Milosz:

Okay.

Ezgi:

obviously.

Milosz:

Right. So it's, it's too simple in principle, and it doesn't, strike people as something that should be impossible. That's a, that's a funny point. Yes. I haven't thought of that. But then for this community events, you alluded to how important is to get people to together. Right. When organizing, this recording, you were in the process of preparing for an EMBO workshop and, uh. You had some interesting takeaways in terms of not only what those workshops do for the workshop participants, but also how, interesting and stimulating they are for the tutors.

Ezgi:

yes, this is a very good point to, um, bring up, I think, and to discuss a bit. so the EMBA course that we organized is actually the, uh, fifth round of the same course, but the content is not always different. So obviously it has been updated over the years to match the needs of the community. so it was first started by Alexander and organized two times in Barcelona. then, we brought it here. One we had to do, virtual. And then two times in person we organized it. Uh, ally so far. and in this, uh, round, we try to give as much weight as we can to ai, emerging AI methods, the focus has been all along, of these courses. biomolecular interaction modeling. at first, and for, for the first for iterations, the central idea was how to use experimental data to predict complexes. But this time it was how to use AI and dynamics information to properly predict, uh, complex structures. We had a great, uh, lineup of speakers and, the other great thing was that EMBO provided us, uh, with enough funding so that we can host, our tutor tours, uh, for the whole duration of the course. but of also, they were really kind enough, to stay with us throughout the week. they are all super busy people, so that was really a privilege for everyone I think. and the idea was like we started with classical methods in protein interaction modeling, then continued with of fold and half of fold, extension tools. then, uh, we, uh, finalize our course with, the, with insights from Dynamics. We even had a short, uh, mercury Dynamics tutorial, and on the last day, we had a whole dedicated, day on, disordered proteins and the modeling, the interactions of disordered pertains. we also have some experimental data content, uh, which was very useful, uh, during this course as well. So I don't want to mention each tutor name one by one so people can, uh, look it up. if they Google for ismir and integrative modeling, I, I think this is the only thing that's gonna pop up. And but having all these great minds together, and having a great selection of students. so we had 25 students and they are really carefully picked, we offered a very, um, nice contemporary content. We had a, a good number of applications. So these sessions were very lively and in the end, in, by the end of some lectures, we found ourselves like having a discussion altogether. the, tutors, uh, were, inputting things and then discussing with students. But in some instances, the discussion was going on between the tutors. and then that, that made me realize, like, uh, which, you know, but it's more solid when you see it with your own eyes. The field is developing so fast that is also an opportunity, to have these people to discuss among each other, or it's a, a great idea or great opportunity, because, since the main development did not come from academia. And the way they made the research is totally different than our type of daily research. I think even. Between, professors or PIs, it's now very important to get together in these events and discuss the limitations or discuss the advantages or discuss how you can make use of X tool for Y purpose, because these things are still not very nicely protocoled, put down somewhere. Or maybe there's somewhere, but there's so many things published. basically lose track of it and then X person doesn't know about Y paper, basically. And so in the end, it was a course for everyone and that gave a really

Milosz:

Hm.

Ezgi:

to all people who stayed with us. And it was very energizing and honoring for us as well. and I think especially, um, now it's even more important to organize things like this, uh, because now it's the time to discuss. Basically now it's the time when we really start to understand the impact and the limits of the tools that have been emerging in the last five years.

Milosz:

Yeah, it was a common theme for many past conversations here. Uh, how much science is a social enterprise in the end, and how much It's hard to do science just in our own labs, talking to. labmates, but not really, uh, to other labs across the world. And that kind of leads me to, to the last question,'cause this is something I already, I think, discussed a bit with Vlad Cojocaru, who is now across, the pond from you in Romania, but exactly about, your positioning in Turkey, which is, I mean, it is great that you are now able to organize such events. Uh, and I will, myself be moving to, to Poland. I'm thinking of the issue of moving to places which are not historically as integrated, um, within the scientific community. And, uh, what is your experience exactly? I mean, overall, how do you see, the history of, moving back to your, to your country of origin and establishing. A kind of highly visible lab there, uh, you know, what are the benefits versus risks and how much more work it is compared to, or less work maybe compared to, staying in, let's say Western Europe or the us.

Ezgi:

so indeed as you are anticipating, there are hardships, but there are things that are easier as well. it's just the context and the system is different and will be very different For example, it'll be different, for you in Poland than, what things have been for me in Turkey. but before specifically coming into the advantages and disadvantages. I have to highlight one thing, um, before I forget about it. I think when you move out of, uh, the center of computational structural biology research. and if you talk about Europe and then you start going more towards the periphery, scientifically, then it becomes really important to be a part of a European network to continue being a part of a European network.

Milosz:

All right.

Ezgi:

really what made the difference for me, like at the start. So I am, I am here already for almost eight years now, the first two years. They were really difficult, like extremely difficult. But before coming to what was difficult at the time, maybe I should mention what changed my life by the end of these two years. And it was the, uh, installation grant that I got through embo. So that network, so the installation grants can be also received in Poland. I'm sure that you're aware of it as well. Uh, in a couple of European countries, people, researchers can't, make use of, for example, in the case of EMBO, uh, EMBO membership of the country, as much as a German scientist would, uh, would do, or a, or a British. After understanding this, they release this different call, which is not possible to get, uh, by a German lab or by a British lab. but, uh, uh, someone from Turkey, we are also a member of Embo because of this. Someone from, uh, Poland, someone from Lithuania, et cetera, uh, so they can get this grant. So through this grant, they support you for five years. The money is for our condition. The money was great, so I built my lab from scratch with the money that they provided. the money is flexible, so you can buy whatever you want, uh, of course, as long as you justify it somehow. But the other thing is that they, as soon as you're in there, they integrate you in a network. Go to meetings, through their meetings, through their support. And it's not gonna be deducted from your own research money. It is something extra. So for five years you have this extra luxurious life, let's say. for someone being in Turkey, you don't think about like, uh, which conference should I go? Or, how am I gonna do it? Because the currency changes right now are so crazy, like all these things. stopped thinking about them for five years. And they also support your lab members. And this is, this has been amazing. That was a real, career booster for me. So that was the first thing. and I think. Having any type of, grant or agreement that will help you to keep your contacts or, or to keep you tight to the main European network would be a great plus, would ease your life much better if you and, uh, some other, scientists if they want to go back to their country of origin. the second thing that was, uh, a booster for me was, uh, being a cusp assessor. I guess that came, thanks to the work that I did in Maine or Central Europe. and they, that gives the message like, as long as you did, good work, sooner or later you're gonna get the benefit of it, even though you didn't do that work in your current place because in your current place. What happened to me was it took a lot of time until I could settle my lab things are either slow or things are different. The way sciences perceive the pace, the pace is totally different. Time flows differently. things are you have the urgency here as well, but the type of the things being urgent is more on the bureaucratic end than on the scientific end. So that's why it's very hard to balance these things, especially if you're a, um. Newcomer because you don't know the system, you don't know, which knows will be okay, but which knows you should avoid to say, you know, because in the end it's a career and you have to take care of yourself as well. And not everything is about science. so these things were, really difficult for me to understand the culture, to understand the system, to steer myself with the, uh, more like, uh, central European mindset in my home country, even though it's my home country, I totally changed my world culture, obviously during the time when I was abroad. So these have been very difficult. but the good things, so these, are the disadvantages, but the advantages that being in a place where people really don't understand and must, what you're doing also makes you kind of nicely invisible. So I can work on anything I want to work on, which is great. I really love it. So there's no

Milosz:

That's a nice take.

Ezgi:

exactly, there's no institutional pressure saying like, okay, now, um, I don't know, concentrates on this type of direct development or something, which can happen actually in central Europe because. Depending on where you are, because there's big money coming in, and when there's big money coming in, then the responsibilities are different as well. And if there's not much money, going on, then you don't have the same level of responsibility too. this made actually our research flexible and helped, helped us adapt to the changing atmosphere, changing landscape in the AlphaFold era, let's say. and this also allowed me to do the type of research that I want to do here.

Milosz:

Right. I was always torn between this like, listen to the advisory board, you know what they want you to do versus.

Ezgi:

Yeah.

Milosz:

What do you really want to do knowing that it's not optimizing maybe your career for maximum impact, but is keeping you excited and keeping you, really willing to, to push the thing that you're after. Yeah, that's an interesting point there. uh, um.

Ezgi:

being not seen that much can be an advantage as long as you can keep your good contacts obvious, and as long as people remember that you are there, and I think for this. I didn't start for that obviously, but it was a, side outcome organizing courses in your home place. I think it's a really great advantage also for the people to come and then to see, to see your students, to see your research institute, to see the environment, to see your culture.

Milosz:

Absolutely. I think that there are great opportunities exactly for her. Engaging more people with, science, especially, you know, I'm thinking of my, my Polish background, but places that don't have that much developed science yet often have this, um, kind of prestige push, right? For showing off that, oh, we actually have people who are doing great things. So if you can be that person, it's even better for.

Ezgi:

Exactly. That's very true. That's very true.

Milosz:

So hopefully there's plenty of space for people who want to continue. There's always this question of like, how many people can remain PIs after their postdocs or PhDs? And, uh, of course the mainstream institutions cannot stretch beyond, uh, their reasonable sizes. So we have to have a broader maybe perspective or discussion on. Um, where people can go and, um, how to organize their lives. That that's, that's why I appreciate your feedback there and, and your thoughts on the matter.

Ezgi:

Yeah. Yeah. I mean, in the end, maybe this is the time when our field is saturated and kind of generated enough. human power to go back and to also spread the knowledge. So this is, I was making more pragmatic comments obviously, because also these are the things that people usually don't discuss about. there's also this other great side is that basically your teaching your own people and they're really in need of this. And this is really, I mean, the kind of satisfaction that you get from this is like. You can't measure it, measure that with something else. I mean, you can't measure it with money or with, I don't know, being a, X person at Y place. So you also, you also have these things as well, like the opportunities that you didn't have. Now you can provide them to your own students. Like how nice is this? but

Milosz:

I agree. Yes.

Ezgi:

the emotional bit, which was important for me. That's why I wanted to come back. but on the other hand, yeah, maybe now the field is saturated as well. Now beyond the people who are taking things from this emotional land beyond these people, maybe more people would need to do, do this as you said. Maybe, maybe they would need to locate in places where they're not even from maybe, I don't know, a Turkish people would. an opportunity in Ponce maybe. Um, maybe in time this even, we will come to this point because the field is really growing and there are more and more scientists being, raised. So

Milosz:

True many things to, to think about for the young people out there planning their careers.

Ezgi:

yeah. Tough

Milosz:

Okay? Yes. But hopefully full of opportunities. And it's nice to be on the, on the rising curve it feels good and it, uh, gives us a purpose in the end. Okay. Ezgi Karaca, thank you for being on the podcast. Thank you for sharing your history and your observations and your ideas with us.

Ezgi:

Thank you, Miłosz. Thanks for the invite. It was great and fun.

Milosz:

It's been a pleasure. Okay, thank you a lot. Have a great day. Bye-bye.

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