 Let me see the, where's the pointer? Oh, here. All right. Okay, I'm a one-in-show from the Fred Hutchinson Cancer Research Center. And for those, so this is Mount Rainier Lake Union, and this is Hutch, that's where I work. For people who do not know me, our lab has used engineered communities because they're very well controlled and mathematical models to understand, for example, the evolution of cooperation cheating. For example, we showed that adaptation to a new environment can allow cooperators to purge cheaters stochastically. Moreover, defectors or cheaters can paradoxically create conditions that rescue cooperation from cheaters. We also studied spatial organization that is how interacting cells, cells that affect each other's fitness, organize themselves spatially. For example, we showed that strong inter-population cooperation between green and red would cause mixing of these two populations. We further showed that spatial self-organization favors cooperation over cheating. In this case, for example, the isolation of the blue cheaters away from the two cooperating populations. But today I want to talk about something that's completely new, that's not published. And because I'm giving the last talk of the first day, and I think some of us are probably still jet lagged, so I have decided to change the style of my talk to a little bit conversational. That is no slide for a while. And I want you to interrupt me with questions because I like that far better. Okay, so I want to open the formal part of my talk by posing a question. So suppose I have a community of six, say, bacterial species. And the community does something interesting. For example, it converts cellulose, an agricultural waste to a useful product, such as an anti-cancer drug. But the activity is very low. And I want you to help me to improve the activity of this community. What would you do? Excuse me, what did you say? What's the temperature? What? Right, the temperature, well, but you can try, I mean, of course, one can try random things to see. I mean, somewhat aren't targetedly trying, right? But the combination is enormous. So one might want to ask what? Keep some culture. Very good points. So I will, that's a very good point. I'll come back to that. So the first question one might, the first step one might want to do is to figure out whether a single species has that activity. Because if a single species can't do it, then we don't need to worry about five other species. And we don't need to worry about losing that species. Just in case that species actually grows slower than five other species. We don't need to worry about losing it as when we do serial passage. So go back to lab, isolate the six species and test them one by one. And if we find that no single species has that activity, and that activity I would define as community function. So that is, so community function is a biochemical activity that is possessed by a community, but not by any member species alone. So now suppose that we find that no single species can do it, so it's community function. So we still want to see what is the minimal subset of communities that can do that, sort of like what Pete said. So you can also do standard dropout experiment, right? You drop, you throw out species when you throw out species two and so on and so forth. And so in the end of the day, you might realize you only need two species to have this function. Then what do you do? Excuse me? Ecosystem what? Network selection. Very good, very good. So one possibility is I would come back to that point. So one possibility is to try to figure out why you need the two species to have this activity. That is how species A affects species B and or how species B affects species A. But this is no trivial task because we know for species that are relatively uncharacterized, each species can release tens if not hundreds of compounds. And then many of these compounds can affect the other species in diverse manners. And so suppose you actually, you in fact go to the laboratory and figure that out, you end up with a haystack of interactions. By interactions, I mean instances where individual autophysiology of another individual. So you have this huge list of interactions. But still you will have to figure out from this haystack the needle interactions, right? Interactions that are actually critical for community function. And then I'll give you a hint of how to genetically manipulate species or alter abiotic environment such that in the end you can modify community function. Alternatively, as that gentleman suggested, you might be able to perform artificial selection experiment. That is you grow up many copies of that community, allow cells to grow up, allow them to accumulate mutations. And then you will select among this population of communities, those communities with high activity and then only allow these to reproduce. So how would you reproduce a community, right? Community is not a cell. So you could, for example, split a community into sub, you know, into sub or baby or newborn communities and allow them to grow and repeat the cycles again and again. Conventional wisdom says you get what you select for. But is that true? I want to show you that it is often not true. But if you're smart about how to do that experiment, you could in fact probably get that to work for you. So going back, but before we dive into details, I want to actually take one step back and introduce first the selection of individual cells and trying to think about how selection on cell level different from community level, right? So I want to really ask the folks on cells and the communities, what's the difference? So by cells I hear actually in the talk I will only focus on asexual microbes. So imagine I have a population of cells, they all express JFP, right? And suppose I want the fastest growing cells. What would I do? I would grow up the culture, take out the sample and subculture them and inoculate them into fresh medium. I would propagate, that is propagate this population again and again. And by definition almost, right? The fastest growing cells will take over. And this is what I call natural selection, right? The survival of the fastest growers. Now imagine that actually what you want are the brightest cells. Because expression of JFP usually inflicts a fitness cost. So those cells, the bright cells tend to be slow growing. So how would you select them? Obviously you cannot use this anymore, right? So instead you would use artificial selection. So what we do that you would grow populations of cells up, you will select the brightest cells using for example flow of solder. And then only allow those cells to grow and that's artificial selection. In order for selection to work, you need three key elements. First is variation and you need selection and you need heredity. So variation, mutations can create variations in genotypes and sphenotypes. Selection in artificial selection experiment is done artificially by flow solder. And the heredity, it's also satisfied in this kind of experiments because bright cells tend to give birth to bright cells saving for those rare events where mutations actually break this heredity. So this kind of artificial selection experiments on cells have been done numerous times with great success. So how might that work for communities of cells from different species? So community selection has been done but not many times, only a few times. And I want to present you with example, a typical example. So in this example, for example, researchers are interested in isolating communities that were degraded industrial pollutant shown here as brown color. So what they did is they went to a pond and they isolated a bunch of microbes and mixed it up and inoculated it into 15 tubes. So these I would call newborn communities. Even though we use newborns to usually meet individuals but here I mean communities. And then after four days of maturation time, the cells would grow up and they would start degrading industrial pollutant. So then I would call the last stage adult communities. So bearing in mind that the maturation period of time was somewhat arbitrarily selected. So when I say adulthood, I don't exactly mean some special physiological states such as stationery phase. So among the 15 adult communities, those experimentalists picked three top performing communities with the lightest color and mixed them. And then allow these to reproduce by splitting them back into 15 newborn communities. And this I call selection experiment. Of course, as a control, they randomly picked three adult communities and then reproduce again by splitting. So this is the control. So out of four control experiments, the pollutant degrading activity increased into and decreased into. In all the four selection experiments, the pollutant degradation activity increased in three and decreased in one. So of course the number is too small to tell, right? This experiment is very interesting but even if it had worked, we would not know the mechanisms. First, we do not know whether pollutant degradation is a community function or whether it can be done just by a single species. Second, which is actually quite key point, which they would not know, we would not know whether the selection acted on species or genotypes of species. And I would argue that this distinction is important because if a selection had worked on species like the enrichment culture kind of experiment, then you would expect probably the community function would improve very fast initially as the right number of, the right types of species are concentrated and the species that are not good for the activity function is discarded. But then because there's no influx, continuous influx of species, we would expect that community function would level off also very quickly. In contrast, if a selection worked on genotypes, we would imagine that mutations would create new genotypes and that might allow community function to continuously improve. And this kind of studies would not reveal that. And the third point, which is also a very critical point is that we do not know whether community level selection is in fact needed. Because you can imagine the microbes might use the pollutant as carbon source. So if they grow faster, they would degrade the pollutant faster. And in that sense, you don't actually need the community selection. You just natural selection would already generate exactly the same result. So we want to understand this process better. And I remember going to an evolutionary biology meeting asking three well-respected evolutionary biologists. I suppose I wanted to experiment like this. I would need to know many details. For example, I would want to know how many communities I have to start with, I have to select from to ask them that question. So the three answers are the following. The first answer was that is a great question. The second answer was tens of communities, just like this one. And the third answer was millions of communities because the more the better. Because very little is known about how to do this type of experiments. So we were sort of forced into doing some theory first, you know, after our lab is mainly experimental lab. So we want to consider this actually from theoretical point of view, hoping that what we learned in theory would help us design better or smarter experiments. By way, I mean Li-Shea, a very talented postdoc in my lab. And of course myself is also involved. So the system we are interested in, first of all, we have no experimental data yet, right? So this is just all these things I will show you are computational results. So the system we are studying is a two species community comprising helper H and the manufacturer M. The manufacturer makes a product which I call P. The helper but not manufacturer can digest cellulose on agricultural waste. And as helper grows on cellulose, it releases byproduct, such as acetate or whatsoever. And this byproduct happens to be the sole carbon source for manufacture. And the manufacturer would devote a fraction of the cellular resource to make product. And Fp here is the most important parameter here. F means the fraction of manufacturer growth. And P is subscript for standing for the product P. And the remaining one minus Fp is then used for manufacture to grow. So these two species are enclosed in artificial boundary, community boundary, such as a microtider well. And in addition, the two species would compete for a shared resource such as nitrogen. So for this community, right? So we know that the helper will just grow, right? Because it has cellulose waste, it has shared resource, it will just grow. The question is that whether two species can coexist. So manufacturer initially cannot grow, right? Because they have to wait for the byproduct. But if when the byproduct has accumulated to high enough level, manufacturer actually catches up in a sense that it actually grows faster than helper, then there's a chance that two species would coexist. And in fact, these kind of communities have been engineered. And then when these two species grow, they will depleted resource and growth will cease and the product will not be made, will cease to be made at the end. And these, in fact, in this kind of communities, the species composition can be stable in the sense that if you start the two species at different ratios, the ratios will converge to a steady state level. So now we're comforting thought that we have a community which has, as I said, has been demonstrated in the laboratory. We have two species community that can stability coexist. Then we can define community function. So here I define community function as a following. We have a newborn community with helper manufacturer and then we give it cellulose in excess always. And then we give it a fixed amount of resource. After maturation time T, the cells will grow and they will consume resource and make product shown here as magenta. So we define community function Pt as a total amount of product P at maturation time T. Of course, there's a lot of thinking going into the choice of maturation time T, right? It's experimentalist. I don't want to maturation time T to be too short, right? If it's too short, it's a waste of this resource. And also just mutate, it's very little chance for mutations to accumulate. At the same time, we also don't want maturation time to be too long. Because if it's too long, the cells will hit stationary phase and it will get into very complex physiological changes as Professor Huang just talked about. So we chose maturation time T such that the majority but not all resource has been consumed. Just because these are the considerations as I as experimentalist would entertain. And now what affects community function? Of course, initial conditions, right? The initial numbers of helpers and manufacturers in the community, the resource. And also there are parameters and by parameters for biology audience, they're just phenotypes of cells that we can actually measure in the laboratory. For example, the growth phenotype of manufacturer includes the maximum growth rate of manufacturer of helpers and helpers affinity for resource. Here we don't consider cellulose because it's present in excess. And for manufacturer, it's maximum growth rate of manufacturer, manufacturers affinity for the byproduct and for the shared resource. And then of course, the manufacturers FP and that's a very critical parameter because that is involved both in growth and in the production of the product. The other parameters that we consider not changing evolutionary, not capable of changing dramatically. For example, the release rate of byproduct as a function of growth rate. And also the conversion factor that converts the resource to cells. So then we have these six changing parameters, right? So it turns out that in our system, improving cell growth, that is you improve the maximum growth rate, you improve the affinity of cells for resources that actually improves community function. In single or in combinations. This doesn't have to be true in general, but it happens to be in our case, it is true. And the parameter range we chose for this model is based on yeast or it based on the body is saccharomycesae reese. So this turns out to be very convenient because that means that we can fix the other five parameters to the biological maximum. And only consider one parameter change FP. So now if we fix the five parameters and ask what FP would give you the maximum community function, we get this. So maximum community function is achieved at the intermediate level of FP. And that's very intuitively, it's very intuitive because if FP is zero, right? So the no product is made. They devote nothing to making product. But if FP is one, they devote all its resource, all the manufacturer devote all the resource to make product, it won't grow. Then it would be outcompeted by helper. So then how might we experimentally realize that optimal state? Bioengineer would say, well, why don't we pre-optimize the helper and the manufacturer individually in the sense that it will make helper grow as fast as it can so that it makes a lot of byproducts. We can also make the manufacturer make a lot of products. So let's try that. Let's go back to the previous slide because as I've just shown you that for helper, if you increase the cell growth, you actually improve community function. So we can just simply do a natural selection experiment on helper, that is we just grow it and do serial pathaging it in the presence of the shared resource. But for manufacturer, it's a lot more tricky, right? Because we cannot do natural selection because natural selection would favor zero FP, non-producers. So one might think, okay, why don't we do artificial selection on manufacturer based on its ability to make this product? So experimentally you can do this, right? And there's a large literature on group selection, that is you select groups of cells of the same species. There's a large literature, so I do not have time to cover it. Suffice to say, from literature, it seems like the best way of doing such experiment is that you put a single M cell, one each, into microtider wells, right? And you give it resource and give it by product. And you let it grow for mature maturation time T and look at which well has the highest amount of product. And you only allow those manufacturers to reproduce. Sure. Oh absolutely, I agree with you, yes. Yes, but in our parameter range, this happens to be the case. It happens to be, that's true, that's a very good point. It doesn't have to be. And so in this case, so we could, but then the question is how much byproduct would you add to a well, right? So you could, ideally you wanted the byproduct dynamics to be like what's happening in the community initially low and as helper goes up, it goes up. But it's very hard to do experimentally. So for simplicity, we'll just add excess amount of byproduct. So it turns out for this kind of group level selection, artificial selection or M, we get that the optimal FP also is at a curse at the intermediate level for exactly the same reason, right? For if you have zero FP, the manufacturers don't make any product. You don't have zero any product. But if FP goes one, the single cell is not dividing and might even die. So you have this low product. But we noticed that the FP, optimal for monoculture is below that's optimal for community. And that actually is not surprising in the sense that in monoculture selection experiment, we're giving it excess byproduct, which is in contrast in community where the byproduct is provided by the helper. So now comes to the first major conclusion of this talk, which is that you could pre-optimize each individual species, but the monoculture optimal may not equal to community optimal. So now the question becomes this, right? So if we start here, the monoculture optimized state, we know that natural selection would produce will decrease FP to zero. So the question is how might we do community level selection such that we will push FP up against natural selection to that which is optimal for community function. So now I will go through the selection process. So we start with newborn communities. Of course, we can start with hundreds or millions. The more communities you have, the more variations you can select from. But also the problem is that the experiment becomes much more expensive. And here we have, we choose 100 communities and each community has about 100 helper plus manufacturer. And we choose number 100 cells so that we don't accidentally lose one species. And the ratio is about one to one, but the ratio doesn't really matter because it doesn't matter which ratio you start from. They converge to the steady state ratio, which is around one to one. And then we give it excess cellulose and affix the amount of resource. And we wait for maturation time T until it's the end of that. And so we, as I told you before, the maturation time T is such that the resource, the majority of the resource is used, but not all of it is used. And we put enough resource such that during this process, the total population will increase by about 100 fold. And here during the maturation, mutations will occur. And I said because we have fixed the other five parameters to the biological maximum. So the mutations can only happen in FP, the fraction of manufacturer growth devoted to making product. So we pick the mutation rate, so mutations will occur with a probability determined by mutation rate. And the mutation rate is the highest that has been observed among the mutators. So it's still relevant, it's still realistic. We do this to speed up our computation. But we have also tried it with a lower mutation rate and the same conclusion holds. Now the phenotype, how would the mutation change the phenotype? We base our parameter choice on a recent study on extensive mutagenesis on JFP. So what they found is that the vast majority of mutations are neutral. And a small fraction of mutations would make a non-mutants that is make FP to zero. And then the rest of mutations will make small changes, right? Plus and minus a few percent of the FP. So then we can, so out of these, out of these communities with highest function, we will allow that to reproduce in a sense that we will split it into newborn communities of about, on average, about 100 total cells. And if there's not enough communities, we'll go to the second highest performer and continue this process until we recover a total of 100 communities with each with about 100 total cells. And this process repeats, the cycle repeats. So let's consider the three key elements here, right? So selection here is very similar to selection on individual cells, it's artificial selection. We pick whichever community that has highest function. So variation here, mutations will introduce variations among individuals and thus variations among communities. But there are additional sources of variations. During reproduction, each newborn community will sample a subset of the genotypes of the parent adult community. As a result, these newborn communities will look different from each other, that's variation. Moreover, even though we are targeting on average 100 total cells, some communities, some newborn communities would happen to have 80 cells and some would happen to have 130 cells. And this would make all these newborn communities look different from each other and also from the newborn of the last cycle. And furthermore, there's another element that might compromise heredity, that is due immaturation. A genotype frequency can change rather rapidly, making those ones look different from the newborn of the last cycle. And that's my definition of heredity, the similarity of newborns from one cycle to the next. So the question is that will heredity be strong enough for community-level selection to work? I'll show you some results. So first is random. Random selection, we randomly pick adult communities to reproduce. And not surprisingly, the community function declined very rapidly to zero. It's because making product has a fitness cost. And in the absence of community selection, the non-producers are favored. Now if we apply community selection, we see this. So the three blue colors are three independent expanse. We see the community functions fluctuate around the starting point marked here by the magenta cross. It's better than the control, because it doesn't go to zero. But these are nowhere close even to a theoretical optimal. Why is that? Why doesn't work? So we looked to see within one selection cycle, how does community function correlate with various properties of newborn communities? The reason we look at the newborn communities is because the mutations occur rarely, and also the frequency, because it's a small change in fitness in FP and fitness, so they actually mutants, even if they arise during maturation. They don't, the frequency does not increase that rapidly. And thus the community function at adulthood is almost completely determined by the newborn state. So first we look whether that's correlated with FP, because that's the thing we want to select on. We see very little correlation, right? So these two dots are the two newborn communities that when reach adulthood being selected, because these two dots have the highest community function. But if you look at the average FP, average among manufacturers, it's actually below average. Now if we look at, if we correlate, if we see that, we see that community function instead is highly correlated with the total population size of newborn, because these two communities, it's identical to these two, happen to have higher number of total cells in the newborn state. And these also have a higher fraction of helpers. So why would this correlation exist? If we look at the dynamics, it becomes clear. Remember I told you that we chose maturation time T such that the majority, but not all resource has been depleted to avoid the stationary phase. So during this maturation time, the cells will grow and the product will be made. And this is for average community. But by chance, some newborn communities could start with higher number of total population, of total cells. And these ones will be able to deplete the resource more thoroughly and make more products. Even though the lucky community has exactly the same Fp as the average community. And for helpers, the fraction helpers exactly the same reason. Because if you happen to sample more helpers, the manufacturers will experience less lag and then they will be able to quickly grow, more quickly grow and deplete resource more thoroughly and make more product. So a prediction from this kind of reasoning is that if we fix the number of helpers and the manufacturers in each newborn community, this will work. The community selection will work. This, in fact, can be done experimentally by sorting the helpers and the manufacturers into each newborn community. And when we try that, when we fix the number of helpers and the manufacturers in newborns, we indeed see that the community selection now works, at least in theory. A second possibility is that, going back to this, is that we could, for example, extend the maturation time just by a little bit, such that those average communities and unlucky communities would have time to catch up to those lucky communities in dashed line. And this would allow all these communities with the same Fp to have exactly the same function. And we tried that. We extended the maturation time 20% longer and that also allowed the community selection to work effectively. Although a cautionary note is that, then we will need to deal with heterogeneity introduced by different duration, different lengths of duration in stationary phase, which might also create experimental complications. So to summarize my talk, I've shown you two points. First, that we might want to pre-optimize member species before embarking on community function because that sometimes can help. Although a monoculture optimal may not equal to community optimal. And the second, we see that large non-heritable variations can arise during reproduction. For example, when newborns form, the numbers of helper's manufacturers can dramatically influence community function. And thus making the selection of Fp much harder. That is, you have much higher noise compared to signal. That is, you have much higher level of non-heritable variations compared to heritable variations, which is mediated by Fp. And these kind of large non-heritable variations should be minimized for selection of community function to work. So with this kind of theoretical framework, we want to ask the future direction to ask how numbers of communities or the total population size in newborn, the mutation rate, and the selection strength, why all these factors would actually affect one of these three aspects. How would these changes in these various experimental parameters affect the rate of improvement of community function? So with that, I want to thank the organizers for putting together this wonderful symposium. And I welcome your questions, thank you. Questions? Yes, Chris. Well, that was fantastic, both in terms of content and in terms of style. It seems, so the second one on your list, the total population size of the newborn, I'm gonna put a vote there. Have you played around with that at all? Because it seems like you've got this inherent conflict if you've got producers with different fraction going to product, those that make more product will be, that's a good group trait, but it's a bad individual trait because they grow more slowly. And so, if you dilute much smaller, and it's okay if occasionally some of your populations even go extinct because they don't have one of the two, but then you could purify those that have a high production of product and have no low product. I see your point, but it's more complex than that. Because if you manufacture, grow, you have more producers, even though each producer is producing less. That's why for the monoculture, it's also intermediate. FPC, you see what I'm saying? Because a single cell just can't do, even if you spend all your effort making product. It's not as, see, as exponential growth of manufacture and each of them spent less in making product. Sorry, I wasn't quite entirely clear. It won't monotonically select for higher and higher because obviously it can't go to 0.99 or whatever. But you should, in very few cycles, be able, whatever that intermediate optimum is, 0.5 or whatever the optimum is, I think you could really quickly purify for that number, whatever number best suits the otherwise corners. In monoculture selection. No, with group level selection, but going to smaller bottlenecks. Smaller bottlenecks, you mean model, I mean, M selection or M, group of M. Selection, but bottleneck much more strongly so that whatever intermediate level of FP is optimal for this group selective trade, I think you could get there rather quickly. More quickly than that. Yeah. That's exactly, you have exactly the right idea, right? Because I didn't have time to cover group level selection, right? Because the newborn size is critical for the rate of evolution, right? The rate of improvement. That's why we want to try total population size in newborn. That's the second item. You're exactly right. That's right. One question here. I'm not sure if this model can be generalized for any product because, you know, like if the product is obtained by a chromosomal gene or a Blasmid gene because the Blasmids are not stable, but chromosomes are stable, so. Yeah, that's a very good point. That's a very good point. Because that's a no-mutant I'm talking about, where you can lose plasmid, right? So the rate, the mutation, we actually try, you know, yeah, try different mutation rate spectrum, right? So if it's a plasmid loss, it's a much higher no-mutant set. We have tried that. That actually does affect us quite dramatically in the sense that whether you can purify away those non-producers from the community. That's a very good point. We have a question here. Can you go back to the slide when you showed P versus T plot? So this is something interesting, the second and the third one. It's like of, it shows some cutoff and in time, and after that it booms, right? That's the time where we harvest the communities and do selection. That's like arbitrary determinant, it's experiments that you have to choose a time at which you will do the selection, right? So this is the longer, sorry, I should have marked, it's the longer maturation time. You extend it by 20%, that's about 20% extension maturation time, such that the unlucky communities, the average communities would catch up with the, I didn't explain that well, most likely, right? So this is average, this is average community. I should have shown one maybe it's of a less lucky community, right? They would grow even less. So then you would have, even if these communities have identical average FP, they would actually reach very different levels. And when you harvest them, if you harvest them at this point, they will look very different. And you might pick this one, even if this FP is smaller than the other one. So that's what happened earlier. So you pick this one, even though this has very low, like a lot of non-producers, but just because it happened to have more cells, it gets picked. And the more cells is not heritable, because the next round you may not have more cells. So if we extend the maturation time just by bit, by little bit, right? So all communities, regardless of the stochastic fluctuation at the beginning, right, they will all in the end reach the same point if they have the same FP. So then if you have higher FP and a higher, you get selected. So that's why once we extended the maturation time T, or not considering the physiological changes associated with stationary phase, right? Then this really helps the community on function selection. Thank you very much. Thank you. Thank you.