 All right. Yeah, hi, I'm Thomas, Gregor. I'm from Princeton. I am thankful for being able to present at this conference. I'm giving you two lectures. And I'll give you just a little bit of my background, because I guess it's good to know where I'm coming from. So if I use strange language, you can interrupt me. Like physics language or biology, that's not good, or vice versa. I felt like the other lectures, you should interrupt me much more than what you did in the other lectures, because I'm much less clear than Frank or CP, so be ready for a messy lecture. So I've come from a few years' background. I started as a theorist. And during my PhD, I realized that, and I was always interested in biology, but I used to do quantum physics. And during my PhD, I realized that, yes, biology is really what I wanted to do. And quantum physics got a little bit boring. And at first, it was natural to switch to a theory lab. And what I realized, though, is that if you do theory and biology, it's very nice to come up with very complicated models. But what's really lacking in biology compared to physics are very good measurements. Because if you want to have a good model that has any predictive power or has anything to contribute, I think it's very important that you, at some point, are able to go beyond the phenomenological description and really try to predict something that can be measured, that has error bars, and where your model really helps you to refine your exploration of that political phenomenon. And so that's what I engaged in. My PhD was in between the physics theory and experiment. And now in my lab, I still kind of do the same, although because it's much harder to do good experiment than biology, physics-style experiments. We are doing mostly experiments. And for the theory, we often actually collaborate with theorists that are in-house or also across the world. And so I'll give you, in the two lectures that I have, a little tour through some of the explorations that I've done. All I'm going to talk about have to do with the prosophala embryo, which you have seen a lot in Frank's lectures already. I give you my perspective on introduction, and then we'll see as we go along. And so I want to also do something special in those two lectures here. I didn't go through my early papers and kind of lined them up and give you lectures about them. I want to give in both lectures a lot of material that is actually unpublished. And so I'll bring you right there. Let's see if that works pedagogically, but it's the first time I'm trying this. And so you kind of have to help me if I'm having too many holes in explaining you things about the background and how we want to do things. And so very importantly, interrupt. So yes, we are interested also in fly development, sorry, in development in general, or in understanding how an organism, when it starts from a single cell and grows into many cells, how do these different cells in that organism during the course of development know where they belong? How does a cell know it has to be in the future head? How does a cell know it has to be in the future toe? It's a very important decision. And typically those decisions are made very early in development. And if you look at these early stages in development, here I'm showing you some four movies that are all stole somewhere from the web. And I didn't mention because I was too late, too short in time, who they are coming from. So maybe some people in the audience don't be offended. Happens to me as well. So this year is a mouse embryo. There's a fish embryo that you probably have seen in CPs work, here's a worm embryo you may see later, and here's a fly embryo. And what all these things have in common is that they start from a cell mass, very few cells. And all of these cells are completely identical. They could, each one of them, you can move them around at that stage in the embryo, and they could take on the fate of any of the other cells. So that means you start with a mass of homogeneous cells. And somehow there are now signals that somehow sometimes come from the exterior, but sometimes often also within that system that kind of specify those cells and determine what those cells will do in the end. And so that is kind of at least of today's lecture and also even of tomorrow's, something that keep in mind that I will mostly focus on. And so we are focusing on the fly's development. And here you see basically 24 hours of a fly embryo. As you saw in Frank's talk, the fly undergoes several stages. At first, the mother lays an egg. That egg takes a day to hatch, as you see here, into a larva. And then that larva takes six days to get bigger and eat food. And then over four days, there's pupation. Frank showed these beautiful movies of how wings, among other things, form, actually all the organs form during a pupation time. And then a complete different organism, a fly, comes out and flies away. So the larva can only walk away after 24 hours. And the fly after 10 days can fly away. And so we are focusing on the very earliest stages. So here's again this picture that Frank has shown. Just looking at this early embryo here. And what we would like to understand is how, when the larva comes out, the different segments of this larva have been formed, which means that you make cells. You make more and more cells. And eventually, a process called segmentation, which is a pattern-forming process, has engaged. And which you see very nicely at the outcome. But there is a blueprint of this already enshrined after three hours when you have cells at the surface of the embryo that express different proteins, different genes. And these genes have a one-to-one connection. See, they're having the same kind of striped pattern here. And there's a one-to-one connection between this gene expression at three hours and this larva that comes out after 24 hours. And so within the first three hours, something really crucial and important happens for the larva to determine its future structures. And it turns out that after three hours, you can, at certain places in the embryo, get individual rows of cells that have a specific morphology. So here you see these cells look all seemingly identical. But there's one row that has, it's called a furrow that has some sort of an indentation. We're going to talk about this one row in a bit and see how sharp this already is. This goes back again to what Frank has been talking about. How do you make these sharp boundaries along an entire axis? And so here in this cartoon, I just want to emphasize again, because I'm going to stop talking about it for the rest then, that after three hours, you have these patterns. And there's a one-to-one connection between those patterns and the larva. And there's a one-to-one connection between the larva and these patterns in the fly and the other fly. So the setup in the first three hours is crucial for this machine that crawls away after 24 hours or that flies away after 20, after 10 days can work. And so that's why we're going to focus on those first three hours. And I'm just going to give you a brief overview, I guess, something that Stefano has alluded to. I just, again, for sake of coherence of my talk. So the fly mother, she lays an egg that's roughly half a millimeter in size, so you can still barely see it with your eye. And that egg has a single nucleus and no walls, no membranes or whatever. So one single membrane that's around the embryo. And then over the course of two hours, that nucleus divides. It just divides, doesn't make any cells. It's called a sensation, which you see here in this movie. So in this movie, we have tagged this nuclei with GFP. Is GFP OK with everybody? Or shall I tell you what GFP means? Show me if you need GFP explained. You want GFP explained, OK. So these nuclei, if you just put this embryo on the microscope, you don't see anything. The thing is white. If you look at it in fluorescence, the thing glows a little bit green. But there's not much that you can see. What you would like to see is, in this case, nuclei. And so what people have figured out, it got a Nobel Prize a while ago, that you can take a piece of DNA from a jellyfish that encodes for protein, that is called green fluorescent protein, that when you shine a certain laser light on it, it will glow. And so you can put this with molecular biology cloning and whatnot into a fly, put it in a germ line so that the next fly kits that come out have it out automatically. And then you can have your undergraduate put it on the microscope and get a movie like that. And so I'm going to use this a lot as GFP. That's why GFP is very important. And GFP tags proteins. And so what I've achieved with this is I put basically a piece of sequence that encodes for this green protein, next to, in this case, a histone protein. Histones are nuclei. And so whenever the fly mother makes a histone protein and puts this in the embryo, this histone protein is now glowing green if you shine laser light on it. And so it's all over the embryo. That's why everything is a little bit glowing. But in the nuclei, it's concentrated, which is why you can see the nuclei. It has more, OK, it has higher contrast. All right, and so back to the work here. So basically, within two hours, you see these nuclear divisions, 13 rounds. They're very beautiful. They're synchronous. You can ask how synchronized they, et cetera. So there's already lots of physics just if you look at these movies here. Then after two hours, miraculously, they stop. So there's a lot of research going into how do they know how to stop. And what happens then over the course of the next hour, there is a cell membrane that comes down from the cortex. So this here is the outer shell. This shell here, we just did a little movie in here. And you see there's a cell wall coming down, OK? Well, cell membrane. I say wall, but it's not a plant. Last time I went to a summer school, they thought I was talking about plants because I use wall. I use the word wall because I'm a physicist and they don't distinguish, OK? It's a membrane. But I might say wall again. And so this process takes one hour. So basically what the embryo has done is, in fact, it's the world champion in making and proliferating cells. Every 10 minutes, bam, bam, bam, you get another factor of two more cells. And it only goes so fast because it doesn't bother making cell membranes. It basically does that in a process later once it has 6,000 cells. And so then after three hours, a process called gastrulation happens. That's what CP has talked about a lot in zebrafish, but it also happens in flies. It happened in you. You did this. Some people say that this was more important. This is the most important event in your life, more important than your marriage or whatnot. And at this level, it becomes a little too complicated for me. I'll pass it on to CP, and I'll just study everything beforehand. So I'm trying to basically set up the system of how do we get to this stage here when everything looks identical. But the cells, of course, all have already individual different identities, such that they can generate the forces that are necessary to gastrulate. So in a way, my work is lead up to CP's work. I want to give him his initial conditions so he can look at gastrulation. Or in fact, in my case, it's for people to do gastrulation in flies. All right, so how do we get there? How is it possible that there's now this one single row of cells here that does something special? You see this row splits in, in fact, the future head of the embryo from the rest of the body? Well, it turns out that at the same time that these nuclei divide and you have all these dynamic processes in parallel simultaneously, you have a cascade of genes that gets transcribed. You have three hours to transcribe genes. And so what happens is that the symmetry is broken by the mother, right? You look at this egg and you see it's not symmetric anymore. So it has poles. And so at these poles, what the mother does over the course of several days, like two to three days in the mother's belly, she puts all sorts of good stuff in the embryo that she's preparing. It's called an oocyte at that point. And among others, it's the histones that you have already seen, because the embryo can't make them itself at that stage. So she puts it in there. But she also localizes sources, they're called mRNA molecules, at the different poles and these sources upon fertilization. They're sources for protein. They get translated at time zero at this point here. They start to be translated. And they establish gradients that span the entire egg. So there's a protein now that gets made here at this source and it diffuses along the embryo axis to establish an exponentially decaying gradient. And that protein is a transcription factor. It enters the nuclei, binds to DNA, and tells downstream genes to turn on or not. And there's several, there's three types of these. There's one that comes from this end. There's another one that comes from that end. And there's a third one that comes from this end, which is something I'm not, I'm only going to focus today about the long axis called the anterior-postia axis. And so then you have a cascade of genes. So there's these broad expression patterns that are directly determined by these so-called maternal gradients. And then these turn on more refined gene expression patterns and seven stripes. And then together, where a row of this guy and a row of this guy overlap, so red and green in this universe gives you yellow, that yellow row determines this indentation. And so you see roughly how this works. That doesn't mean that this explains why this is a single row of cells and why this row of cells is here and not here. And so that's something we want to focus on in a bit. All right, so to recapitulate, because I'm going to use this continuously, the mother puts up a source of mRNA molecules at one pole. That's why it's called maternal mRNA. That mRNA, at time t equals 0, starts to be translated, makes protein that diffuses into a protein gradient. That's why it's called maternal gradient, because it stems from the mother. And then there are several of them. And these gradients turn on a first level of genes. They're called gab genes. So those are zygotic genes. So they're genes that now the little embryo does itself. It doesn't need the mother for that. It just needs the mother to instruct it where to turn on which one of them, or the mother signal, at least. And so then together, these guys, those maternal gradients plus those broad gab gene expression patterns, they determine these more refined so-called peryl genes, which come in seven stripes, in which you may have seen at the cover of magazines or winter schools. And then these individual stripes, they overlap. There are seven different genes of this class. There are seven genes that make seven stripes, if that's the coincidence or not, I actually don't know. But it's seven and seven. And so where they overlap, they can determine a single row of cells that have the first morphologic change in the flight development, if you want. All right, so here are, again, the times. So in over two days, the mother prepares this. And then over three days, it takes three hours to get to the first morphological marker. And the network, if you want, the gene regulatory network, has three different levels, at least. That's what I want to focus on. There's these maternal inputs that the mother sets up that get inactuated at time t equals 0. There is this processing layer, which are called gap genes. They're coming broad expression patterns. And they interact with each other. They interact with the maternal. And then as an output, you see these seven striped genes. And there are seven kinds of them. And in order to identify, though, individual cells, it takes another three to five hours. At least that's the current picture. Because there's yet a fourth layer of cells. They're called second polarity genes, which come in seven and 14 stripes. And so it's believed that over the course of these three to four hours more, on top of those three hours, that every single column of cells here will be able to determine its fate. That's the paradigm. And so of course, in this system now, there's many things you can ask. One thing to ask, for instance, is how precise can you position a single row of cells? Why is it just a single row? And why is it here? And not five rows over or one row over? If you look at many embryos and you line them all up, if that single row here always at the same location, or does it fluctuate a bit? Because if it fluctuates a bit, there's error, and somehow the system has to cope with that in the end. Or maybe not. So these things, we would like to figure out. Something I took out because I saw that Frank talked about it in depth. In this system, we also have to see the problem of scaling. Because I told you that things happen by diffusion. So a gradient is set up by diffusion. But it turns out that these embryos that come in very different flavors and sizes. So there's a factor of five indifference between different species that all have enshrined that same network, that same patterning network. And so they all make the same kind of patterns with the same genes in the same amount of time. But in one case, you need to diffuse five times more than in the other. So how do you do that? Or is diffusion even a valid mechanism for this to work? Because we know from physical pattern formation, which works by diffusion, if you change the side of the box by a factor of five, you make five times more patterns. The wavelength stays the same. Take icicles in the winter. They all make the street lamp twice as large. You have twice as many icicles. And that's a equilibrium between tension and gravity. So if there's something special in the biological pattern formation, because obviously that wouldn't work. You don't want to have a fly that you make five times larger and it has five times as many heads or pairs of wings. So in biology, something is controlled in a different way. I'm not going to talk about that today. Let's go back to the question about precision in development. So you generate a fly in 10 days. And somehow, in the end, things look quite reproducible. If you look at the wing that I've put up here and the wing that Frank had this morning, they look very similar. We know another developmental system, which are humans, of course. We know that twins in humans, they look very much the same. Sometimes you can barely distinguish them. So we know that there's a potential for development or developmental machines, programs, to produce something that looks very much identical. And in fact, the wing of a fly should be reproducible, because if it isn't, you need to find a mechanism of how the fly avoids flying against that column or a tree in nature. So there is an incentive for things to be reproducible. But the question is, how do you actually get there? Because, oh, sorry. I just, basically, before I'm talking about this, let me tell you how reproducible things are. So I basically alluded to the precision of a single row of cells in the early embryo, this morphologic furrow that drives in. But we wanted to see if other developmental features are just as reproducible. And so, yes, by eye I can say my wing looks similar to Frank's wing, but it's better to measure. And so let's take a bunch of these wings and morph them on top of each other. So we scale them, we rotate them, and we translate them so that they look as close as possible than they can get. And we do this with 82 of them. And then we ask, where are these crossing points here? You can ask a computer to do that for you. And if you take these 82 wings and put the crossing points all on top of each other, you see that the distribution function of those crossing points is extremely tight. What does extremely tight mean? Well, this is a qualitative statement. You need to compare this to something that's actually relevant to the system. And what's relevant to the system here is the size of a single cell, because it's that size that determines whether you are in that wing vein here or whether you're in the wing blade. And you see that the size of a single cell is largest at the level of 2 sigma of these distributions here. And so those are wings that come from a random collection of flies that I got in my garden in New Jersey. So you take 82 flies that come from my garden. You morph their fly wings on top of each other. And the precision with which these crossing points here are made, you can make this more. This was a high school student who did this. You can make this more sophisticated and actually take the entire structure of these wing blades and whatnot. But just to make it simple, let's just take those crossing points. They are reproducible to within half a cell width. Yeah, you had a question? A little bit of both. They have bigger cells, but they also have more. But the precision is still at the level of a single cell. And so if I take, sorry, if I misspoke, if I take flies from my garden, they're wild type, they have a precision of a single cell. But now if I make identical twins of flies, which is very easy because you just inbrew them so much that you get rid of all of genetic fluctuations, that width of these distribution here goes down from a linear dimension of a single cell to half a cell. And that precision of half a cell is the same than the left and the right wing of an individual fly. So if you give me two wings of a population of identical twin flies, I will not be able to tell you whether those two wings came from a single individual or from two random brothers and sisters, which is quite remarkable. I'm not going to do much more science with this, but I just want to throw this out there. What a biological system is actually able to do, just characterizing this as of an interesting feat. Because in a way, it gives you a very intuitive mechanism to generate symmetry. Because all you need is to make sure that things come out reproducibly in the end. If you start from these wings, in fact, Frank was talking about wing discs, which are lots of cells. But a single wing disc comes from nine cells in a 10-hour-old embryo. So every single wing in any single droophile fly in the world comes from nine cells. And so you can think of mechanisms how an early embryo can count until nine. And if nine on the left side and nine on the right side. And once you accept that, all you need is high fidelity in your process of generating those two wings. And you have symmetry for free. It's a nice way to get symmetry. Which does not mean that if you fud things up, that you cannot still get symmetry. So the other mechanism that compensates because it's such an important thing that the fly doesn't go against the tree, but has symmetric wings. But in principle, they're not needed if everything goes straight. You had a question there? That is true. There is probably error in the production line. But if you look at whatever, the Fiat production line, you get also only cars that end up being cars, but you will never be able to get them as reproducible as this. Another example I typically give, because it's a little bit more complex than the Fiat production line, is if we were to shoot the mast rover for a second time up there, it would certainly not unfold and do its spiel the same way I did it the first time. So getting something reproducible in engineering is a huge problem. And so somehow, biology has figured it out. Yes, it kills a lot of things along the way. But still, of the 82 wild-type flies from my garden in New Jersey, wild-type still has a survival rate of 98% or something. It's very, very high. But those 2%, yes, who knows what goes wrong there? And I'm not looking at that. There was another question? No. All right, so in the end, everything is reproducible. The question is, how do we get there? And so if you take the different stages of development as the different nodes, different stages that you undergo while you develop, one way of getting something reproducible is you start with something that's very noisy, where you have patterns that fluctuate all over the place. And then at each step, you have a mechanism that kind of refines that, and it reduces the error bar. It's like the Wrenben is long. It's a big process, and you can just think of mechanisms that would reduce your noise. And in the end, using noise filtering and error correction, you get somehow to a reproducible outcome. On the opposite side, you have something that is very reproducible at the very beginning, a precise setup. And then what the system does is from going from step to step, from level to level along these different developmental stages, what you do is you try to maintain that level of precision. Try not to lose it. That's a different strategy. And so the question is, what does our system do? And this is something I want to talk about a little bit today. But before, I just want to recapitulate of what I've done so far. So I introduced the system as a pattern forming system. I want to understand how do you make these different patterns, how do the individual cells know where they are and what to become, in what time frame, with what kind of signal, et cetera. But, and it has been classically used as a pattern formation system. But there's this network of the different maternal gap and parallel genes also gives you very nice access to study properties of gene regulatory networks in general. So I can hijack this embryo and use it as a laboratory and study genetic networks. At this all the while, the outcome of the genetic network is significant because it does something for pattern formation. But I just want to emphasize that these are two different approaches of questions that you can have. One is like scaling of pattern. That's a pattern formation question. But the other is, how does a genetic network actually encode scaling? That's a genetic network question. And so we are kind of today, I'm talking about things that are at the brink between those two disciplines that can both be addressed with fly embryo. And tomorrow I'm talking a little bit about the mechanisms of transcription regulation because, of course, this network here is a transcription network. Every single member of this network is a transcription factor. And you see all these errors here. They all influence each other. So transcription seems to be the crucial process that makes this network, brings this network to life. And so understanding something at the molecular level of transcription will then help you to understand something at the network level will then help you to understand the larger macroscopic pattern formation processes. Yeah? Yeah? Sorry, what about hunchback and dry end? I have no idea. That may be true. I have actually no idea. So I used this picture just because it's complicated and nice. I really don't know who influences who. All I know is that there is a feed forward from this level to this level to that level. There's a little bit interaction. And that's all I really want to know. Because I would like my approaches to get me this diagram out. And I'll show you a little bit how I'm trying to approach this. It is true that each one of these errors here probably a PhD student had to figure out. But so now we would like to step back a little bit from this and try to see what is it that really makes this network work in a more global sense. Like how do you get scaling to work in this network? How do you get reproducibility to work in this network? And for that, the hope is that the individual errors, one individual error at least, one particular individual error should not be important. All right, so how do you make a pattern? So the classic model of how you make a pattern is essentially by a so-called threshold readout. And that goes back to the late 60s when it was proposed that the way how you could make a boundary is you take a gradient and you say, well, wherever that gradient is higher than a certain concentration, I take a downstream gene and turn it on. And wherever it's lower than a certain concentration, I keep that downstream gene off. So that means there's something very crucial happen in this nuclei here because they have to determine between on and off. And because we're talking about molecules that are being read out, right? So different factors are protein molecules. And in these early stages, the concentrations are very low. You have maybe tens, maybe hundreds of these protein molecules. And so that means that if your DNA needs to read what is the concentration of these proteins, but we're talking double-digit numbers, your fluctuations of that measurement that the DNA must make in order to read out that concentration, they're huge. And so there's a potential for huge fluctuations of this boundary because any of these cells might be just as well determining we are. If you look at the downstream gene, though, things seem to work beautifully. So this is a gene called bicoid. It doesn't really matter. It's the one that was anchored here and there's a fusion gradient that got established over two hours. And then at stage two hours, if you measure, you see that there is a really seemingly on off read out of that protein concentration. And so the model that has been proposed is that these morphogens, that's what these things are called, they can basically determine at different levels of concentration different levels of different downstream genes. It's called the French flag model and goes back to Louis Wolpert. Now there's problems with this model and one I just alluded to. The number of molecules are very low and so they have potential for a lot of noise. But it's actually very hard to pinpoint that problem. And so, however, in the last decade or so, there are a bunch of labs who have contributed to kind of putting doubts on this model. And I'm not going to talk about all of them. I'm only going to talk about the ones that come from my lab, obviously, because that's not really what I want to talk about. This is not what the talk is about. I just want to give you an example of what you can do. And so it turns out if you take this gradient here and you ramp it up and down, you can move this so-called cephalic furrow back and forth. You basically can make a fly that has twice as much of this source molecules. And so it will create a gradient that comes from twice as high and goes down. And so now you move all of the patterns towards the posterior and it means that you move the cephalic furrow, which is the readout of all of these patterns, towards the posterior. And you can do the opposite. You can ramp this thing down and move everything towards the anterior. And so if you measure now the dosage of how much protein you have and the location of where the cephalic furrow is, you get data that looks like this here. So these different colors here correspond to different fly lines that have different nominal numbers of copies of mRNA, copies of the gene, basically, of PCOR, and hence copies of mRNA in the fly. And so on this axis here, you basically ramp up the source strengths. And on this axis here, you basically measure where is this location of the cephalic furrow. And because this is a log normal plot, if we are in a scenario where this gradient is a diffusion gradient, where you can fit it with a diffusion, sorry, where this is an exponential gradient, in that case, you are expecting your data to lay on a straight line. And because you can measure, so this is your data for these exponential gradients, you can fit an exponential very nicely and you get a straight line. And because it's a measurement, you have an error bar, which is this dotted line here. And this line measures basically the length constant lambda that gives you the sharpness of this gradient. It's e to the minus x over lambda. So now our data that we collected also does lay on a straight line, but the straight line doesn't have a slope of lambda, it has a slope of lambda over 2. And so something fishy is going on here. We kind of do predict that we have the desired relationship, but quantitatively the number is reduced by a factor of 2. And so it turns out that you can explain this at least partially by looking at mutant flies that lack the other maternal inputs. So I told you there's another source here that creates a gradient that goes this way. And then there's a third gradient that has source molecules on both ends, but where the gradient is much sharper, it falls off after like five cells or so. So there's three maternal gradients that span this axis here. And if you knock either one or both of the other two gradients out, you kind of get data points that will start to fall on the line with slope lambda. And so what that means is that the direct readout, if you just have this guy to read out, might work in a threshold dependent manner. But if you have other genetic components that are integrated in the system, there's a network effect that kind of has to be taken into account. And indeed this only works very early. So this only works like at two hours of development. If you look at the same thing after three hours when the GAAP gene and parallel gene network has come to fruition because it's during that hour that this network really is inactuated, things kind of shift from this line to that line. And so there's a dynamic effect that comes from the integration of the different components of the network. And so there's a network effect at that place here and that's kind of what we are after. That's kind of what I'm after in this lecture. How do we get access to those network effects? Any questions? No, I don't think this talks about diffusion because this is really about the readout of the gradient. How the gradient is set up. All I assumed here is that this is an exponential decay and that I can do with a fit and the fit is very good. Now can I explain what this lambda here is? That's a different question. And there comes diffusion into play. But whether I can fit an exponential or not, I think that's independent of that. It tells us something about how this is readout because my readout here is really the output of the system. It's just a phallic furrow, right? It's not really the setup of the gradient. Yeah? Yeah, so you make an embryo and here you put source molecules. It's like in your bathtub. You have a bathtub and you have ink on this side, a blue ink on that side and a red ink on that side and at time zero you put a drop of ink in there. Oh, sorry, no. No, no, no, no, no, no, sorry. Maybe I didn't say this, but the mRNA molecules that are anchored, they only become active when the fly, when the egg gets activated, when the egg gets fertilized. This happens at the same time when the mother pushes the egg out of its overduct. Okay? Exactly, there are frozen patterns. In fact, there is a pH change and the pH is such that when the sperm is not yet in the egg, the mRNA is not translated and as soon as the sperm is in the egg, the pH changes and now the mRNA gets translated. And that starts at time t equals zero of our counting here. And then it takes 90 minutes for the gradient to stabilize. And then once that has happened, this thing MBT starts, which Stefano probably talked about. Did you? A little bit. A little bit. Where zygotic genes get activated and these gap genes, the first readouts of these gradients, are part of that. They start to get activated roughly, you know, in 20, 100 minutes. But really, when they really come to full fruition, it's between our two and our three and I'm going to show you that in a couple of slides. Okay? Yeah, you're right. It's very important to catch the set up. Are there any other questions here? Eventually, the slides that you eventually get out from the presentation, it's a portion like a wild-type slide. Yes. So the remaining correction mechanism, like the laser viewing development, still controls your third rule. So does it happen at the level of the third rule being at your seven stripes or that's a later? Yeah, so Stefano is asking about the question, well, if you make, if you put so much of that source molecule in there, that you shift everything to the posterior. Now all your patterns, all your seven stripes, they are maybe in the, in one third, in the posterior, one third of the embryo. But somehow, in those kind of flies, 18% of those flies actually do survive. So much, much less than the 98 that we had for real wild-type. And so that means that somehow the thing can compensate and there's in fact a compensation mechanism that goes over the entire 24 hours of the embryo's development. Some of it is corrected here and in part this is this correction there, but not all of it. All right, so we would like to understand these kind of network effects. Oh, was that also part of your question? That's what I thought, because how can multiple gradients affect the same gene? Does anybody want to know this? How can multiple gradients can affect the gene? So it turns out if you make, if you make, if you want to make a gene, and I'm going to talk about this more tomorrow, there's a part of your genome that encodes for the part of, that's eventually going to be a protein, but there's also part of your genome to which transcription factors these proteins bind and tell the machinery that makes genes to make their gene or not. Otherwise, each cell would make all of your genes. And those little, those regions they're so-called enhancer regions. And to these enhancers can bind multiple transcription factors. Okay, so you have a piece of DNA that encodes for the future protein. You have a thing that's called a promoter where the machinery binds that makes that protein. Well, it first makes them an A, but never mind. And then around that promoter region there are little sequences to which the proteins, the transcription factors bind and tell the machinery to make or not a gene. Okay, and so those many transcription factors they all get integrated and then the machinery decides, okay, I like that integration and I'm going to make the gene or not. All right, so, okay, so basically if you want to get access to network level effects and you come from a physics background, it turns out that one viable approach in doing this is to try to measure fluctuations in the expression levels of the genes that determine those network level effects. Yeah, in physics, when we measure fluctuations we can measure correlations of these fluctuations across space and time. And those correlations give you often very large-scale effects and there, in my case, would be network effects. Okay, because it's across the entire network that I would like to see which gene influences which and I would like to do that by measuring the fluctuations, by extracting that from the fluctuations. And that goes back to your question, to earlier question before. That's why in part I'm not so interested in the actual links, whether this one, this gene affects this one or this one represses that one, because I would like my approach to back that out. And so, however, if you really want to understand fluctuations of gene expression levels, you need to be able to make measurements that have access to those fluctuations. You cannot just measure the means. And so, in order to do that, we developed a protocol and took us roughly 12 years to be able to do that. So even in biology, you can spend on the same stupid method 12 years in order to get it to the point where you can actually make a physics-style measurement with it. And that measurement, in this experiment, one of the earliest experiments that I actually did when I went into the system, it's based on antibody staining. So you may ask, well, why don't you keep the thing alive? Isn't that a much better measurement? And that's the same question at front I got earlier today. And there's something to say for, yes, doing things live. You don't look at the system that you were meant to look at. You're looking at some sort of a transform at an X-man, right? Because you have now put tags into your life system because you want to measure something. And so it's no longer the actual system you want to look at. So whatever you do, you would perturb the system. And so my perturbation, at least for today, all I'm going to do tomorrow is life. So don't worry, there will be life. But today, the perturbation I chose today is death. Tomorrow, the perturbation I'm going to choose is I put some weird stuff in the fly and it's a mutant or it's a transformer. Today, the perturbation is death. But it's controlled death. And so it took me 12 years to control the bugger's death. Or at least to control the death such that I can still make measurements. And I just want to lead you a little bit through this because it was a tour de force and it's important, oh god, I already talked almost for an hour. Well, who knows? Maybe tomorrow you're not going to see life stuff. So shall I go faster or is the speed okay? Am I boring you guys? Okay, I'm going at that speed then you may not see a lot of life stuff. Pardon? Okay. And so what we choose, so the first thing we did let's choose this layer and try to make measurements in this layer. So we see this as an input into our network and this is an output of our network. And so let's try to see if we can understand those four genes. And in fact, in that layer there may be even more genes than four but those are the four major ones and there's reasons to believe why they are and we can't talk about those later. So let's just see if by just measuring four genes how far can we get. But if you want to measure fluctuations of genes you need to make measurements of these nodes, simultaneous measurements of all nodes, right? Because otherwise you measure one gene in one embryo and another in another embryo, well you have no idea of how they were correlated, right? And so that's already not trivial because the technology as I said is antibody staining works is you take a piece of the protein that you want to stain and you inject that piece into a rabbit and that rabbit is going to make antibodies against that piece of foreign protein. You bleed the rabbit, you get the antibodies and now it's a rabbit antibody and you can buy from Sigma a little general class antibody that will detect that antibody and you can fluorescently label that one. Right? And I pretend to be able to make that quantitative. Just, you know, just let it sink in what a weird way this is to make a measurement. You take you take a thing that you want to measure put in a different animal, have that animal's immune system react to it bleed it, extract the antibodies take your embryos, kill them with methanol, with a microwave, it's all sorts of stuff you poke holes in them nothing goes in there. Once the holes are in there you take your first antibody from the rabbit and you put it in there so that antibody now hopefully detects a few of your proteins now you wash them away because you want there's lots of excess antibodies that you don't want to be recognized then you take the second antibody that hopefully detects only the antibody from the rabbit so that works by taking a bunch of proteins from rabbits and inject them into goat and have the goat immune system make rabbit antibodies and those sigma labels with you with Alexa probes of your chosen color and then you label it then you apply that to your whole embryo you wash it free again and then you do this four times because you want four genes and for instance a mouse and a rabbit sorry you do this sorry I already used a goat and a rabbit right but now I need to use six more animals because I can't use again a rabbit because then that antibody would detect the wrong protein or would be confused which one to detect and so you need to use lots of animals and they also should not overlap if you use a mouse and a rat they're very similar and that doesn't work and that's just of the antibodies of your spectra right if you want to use red and infrared and yellow and green you need four different colors maybe you want to also see your nuclei that's a fifth color and you won't want those colors to overlap because you want to make measurements then you run into trouble and so let me tell you that these measurements these problems are easy that's not the biggest source of mistake the biggest source of mistake comes from the fact that these are your measurements when you're done I just want to give you so we identified I think 12 sources of error there's error that is systematic and there's error that is measurement all the stuff that I talked about today just now was systematic error so there are things you can do it just takes you a few years to figure them out systematic error are very hard but once you're done with systematic error you still have measurement error and so I'm going to give you a few examples for systematic error and a few examples for measurement error just to see how we deal with them so here's the first example for systematic error you take 170 embryos and ask your graduate student to measure this is just for one gene and so what they do is they kill embryos and put them in a microwave and it freezes development at one stage if you get a good grade sample of flies you can actually work that sample of flies such that you only get embryos between hour two and hour three because that's the time frame you want to look at and that's an easy time frame because I told you there's no nuclear divisions and so I can just count nuclei and I know oh I'm in that time frame but then you put this do this with 170 embryos and you see this mess and so part of this part of this mess comes from the fact that there's time still a one hour window and over that time the gene expression gets really turned on and evolves and so you need to kind of deal with that and so you need a better measure of time than just counting nuclei because with counting nuclei you just eliminate the embryos that aren't even on this slide here and so you need to find a clock and one way to clock the embryo during that time window during that one hour is that membrane that comes down from the cortex because you can measure the progression of this membrane as a function of time and you see that it's a monotonically increasing function right? that's good because now I can use this distance and read off time I just use this function but that only works of course if the embryo in which I measured this which was of course a live measurement and the embryos that are on my slide where I can also measure this here they are reproducible I need to figure out that's actually a valid way to look at it and so here we repeated this measurement this live measurement in 10 embryos and you see they all fall on top of each other there are error bars here these black things are error bars and they can barely see them so much for developmental reproducibility so those are 10 different embryos they have nothing to do with each other they are not even identical twins they are just wild type embryos they are indistinguishable along this curve here and so that means first of all that we have now a way to estimate time with 2 and 1 minute precision depending on where we are what the slope here is but it also is a nice example beautiful example for how reproducible things are at that stage of development and so once you do this you get a much tighter distribution of expression level so what we are looking at here is the intensity along this rim here so you take a little window you slide it along the rim and you have an intensity as a function of egg lengths that's the pattern and that's for 17 different embryos so another source is that these embryos when you throw them onto your glass slide they are round so they have a random orientation when they fall if you really had a random orientation then you would have an issue because the patterns are not symmetric in the azimuthal angle so you need to find out a way to do that so you can pinch them on a little copper wire and do like a chicken rotisserie a bunch of things that you need to do in order to really get rid of noise and what not, of systematic noise and once you are done with your systematic noise you still have noise in the system and that's measurement noise so here I show you for the four different genes the four different patterns in a four minute window of development around 40 minutes plus minus two minutes and you see there is still some residual jitter and here is one profile and it's this one here the bold one and so then you need to go in and quantify your measurement error so your instrument has noise whatever the overlap of the the spectra, there is a little bit of noise the cross reactivity of your antibodies there is a little bit of all of these things where I tried with systematics to take out as much noise as possible there is still some residual measurement noise and when you get all the measurement noise sources together you see so in green in color you see the total noise after we took out all the systematics and in the different shades of gray and black here you see the contributions from measurement error and you see that at no stage do we have more than 20% of these profile variations that are from measurement noise and so the overwhelming amount of noise that we have here is really biological noise and so what I will do now working on is I keep it like this and I will not subtract the contributions of measurement noise and keep that in mind for my assertions that I make about my system because just subtracting the variance is a dangerous thing to do especially if you have several sources of measurement noise because you know they are not all disjoint and so it's better to actually keep make sure that your actual noise your actual measurement noise is very low that your biological noise is overwhelming and then keep working with that with that noise so for each of those sources for each source of measurement noise you need to make an individual measurement and in particular you often need to invent it's always a different experiment and for each one you need to come up with a different experiment and so for instance what the spectral overlap is you can just swap your channels you can swap your antibodies you can swap your colors you can use different colors you can make sure that you make your colors as disjoint as possible to see that your overlap is reduced etc so each one of those sources after the talk we can go through each one of them but there's a bunch and for each one you need a different experiment so as you may have noticed through my talk my premise is to convince you that things are very precise and if I keep my measurement noise in the actual noise if I would be able to take out the measurement noise it would be even more precise you see what I mean so that's the line of work I have to do you keep it you keep it such that it's to your disfavor if I was going to convince you that it's noisy then I can't do that I only can do that because I'm trying to convince you that it's precise yeah very good question I didn't say that the nuclei the cells are only on the cortex those are the ones that give rise to those patterns there are a few that have fallen inside but they are for us boring they're also useful but I'm not going to the only ones that have that matter are really there's a single layer a single shell of nuclei alright so the first thing you can do if you have this you can look at the means and reconstruct the dynamics and so here you see in this one-hour window how the network turns on because you now have a one-minute precision with which we can order the embryos and so this is a reconstructed movie from reproducible time-ordered embryos and so here are the gub genes and here are the parallel genes and you see how they turn on then I can ask well how much do these boundaries here fluctuate and it turns out they fluctuate so you basically measure you look at all the embryos in that time window and you measure the delta X and plot that as a function of X and if you do this at the different boundaries you see in the background the genes at this time window you do this for different boundaries you see that the sigmas hover around the 1% roughly 0.8% egg length line and lo and behold that corresponds to half a cell diameter something that sounds familiar from the wing right but now I wanted to sell you this as a network effect and also I would like to measure not just at the boundaries but everywhere what's the precision of my network of my patterns and so for this this information is in the covariance matrix because the covariance matrix really gives you the fluctuations that two genes have as an effect on each other okay so this is a 4x4 matrix and we can measure that matrix because we have so little measurement there and so from that we can then construct a function that we call sigma X which is a function that depends on locally on where you are in the embryo that basically tells you the combined effect of the fluctuations of all four genes at one given location along the embryo axis you get these derivatives because you do error propagation because what you really want sorry what you did here was you really measure the fluctuation in X of those boundaries right but that's not what the embryo does what the embryo does it experience fluctuations in concentration so in the Y axis and it needs to determine based on those fluctuations not based on them but despite those fluctuations so we really need to take into account the fluctuations in C translate them in X in order to really see how precise the system is and so that's what this function here does and once you're done you see that this is a continuous function because as I said it depends on X and it havers very nicely also around this line of half a cell diameter okay so and of course this could never have been predicted just looking at the means this is looking at the single gene this is truly a network effect it comes from the fluctuations that are in your covariant matrix sorry what they are just nuclear yes the cell size is the difference between it's half it's a distance between two nuclear let's say they will become cells and that quantity is conserved it just shifts by half a cell the time information so I'm only looking at one time window okay I could look at all of them I look at one particular time window and that's the time window I choose that astutely to be the time window when the genes have maximal expression you saw in that movie before they came up and went a little bit down so it's at that maximum peak alright so I'm a little bit torn maybe I should skip something here because otherwise I'll not so I was going to tell you now that there's nothing special about transcription because transcription is a noisy process I told you about this individual molecule that have to be measured etc so that is noisy and there's a lot of noise about transcription in the last two decades that came up systems biology etc and so the question is is there something special about development that things are not noisy is there something special about that fundamental transcription that makes it less noisy in development than in in real in other systems like bacteria or yeast where people typically look at noise in transcription and the answer is no it's just as noise but I think I'm going to do this tomorrow because tomorrow I'm going to talk about transcription anyways and so now I want to tell you just a little bit more about the precision and how we then use this to do more network stuff and so what I just showed you was that we have fluctuations in what I just showed you is that the fluctuations in these patterns have around roughly 1% and that corresponds to this error propagation to 10% in fluctuations in the concentrations okay so the cell has to face a 10% concentration fluctuation but that is enough to determine boundaries with the precision of 1% now it turns out that that precision of 10% is along this entire cascade so if you measure the precision of reproducibility if you want of the number of source molecules from one embryo to the next we developed a method to count individual molecules of mRNA in individual embryos turns out there is roughly 900,000 mRNA molecules and we were able to measure 900,000 molecules with error bars that are low enough to estimate that the flies error is roughly 8% then the protein levels are 10% in the maternals as they are in the GAP genes as what I just showed you and also in the parrol genes and then error propagation tells us that 10% corresponds to 1% and that's what we see in the boundaries along this entire cascade all the way to the cephalic flow okay and so that means that we are not in an error correction type funnel of development but that we are in a scenario where we are basically starting with something very precise and making sure that at each step we maintain well we the system maintains that precision okay and curiously it turns out that it's at the maximal level of precision that it can possibly be at because there is no need for the system to do any better than the lattice spacing which is a single cell right and so there is something optimized here something smells like evolution has driven the system to working as good as it gets that's the first signature actually the flatness of this line here that line here is another signature because that's a signature for information that's processed through this GAPG network is distributed optimally along this along the entire axis okay the fact that the fact that this line here is flat and uniform is a signature that information transmission so information in the Shannon sense of information theory that that information is optimized and you can show that analytically and I was hoping to get that towards the end but I only have 15 minutes left so I doubt I will get there alright so so far it's nice to show that this is what the system has that there is this 1% precision and that it's seemingly optimized etc and you know physicists get excited about this because we are things are precise and it's always exciting to businesses but that does not mean that the biology actually is excited about it and by the biology I mean the system it does not mean that the system actually has access to this 1% we make measurements fine but does the system care about 1% precision and so that's what I wanted to really tell you over the next over the last 15 minutes or so and the first step towards this is to see whether we can make the notion of positional information along this axis here positional information meaning information of the cell about its location whether we can make that a little bit more explicit by using the concentrations of the GAB genes for decoding of the network what I mean by this is whether we can build a dictionary that allows us to take 4 concentrations of the 4 GAB genes and then determine based on those 4 concentrations where is the cell that measures that 4 concentration those 4 concentrations it's a slightly different way of looking at the same data alright so how do I do this so what it means is I need to construct a probability distribution that takes in the 4 GAB gene concentrations and spits out a probability that tells me at what position am I likely to be given that I have those 4 concentrations okay that's a code I'm building a code book and so what I measure however is the opposite I measure the prior if you want I measure given a certain position along the axis what is the probability of finding a certain concentration because those are these profiles that I have been showing you for the last hour okay so here you have concentration as a function of position but if you do this for many embryos you can construct a probability distribution of finding concentration along a certain position along the embryo okay and so we have access to this guy but we want that guy and so we use something called in order to transfer it and so Bayes rule is a very popular rule in statistics that essentially transfers exactly this function to that one by this transformation here okay where P of cripple P of K I sent just for one of those 4 genes it's called cripple P of cripple concentration is a projection of this probability distribution on the y-axis so this is P of cripple here and P of X is the probability of finding a nucleus somewhere and you can just assume that that's uniform because the nuclei are uniformly distributed along the surface of the embryo okay there's no denser regions or less denser regions at least not at the level at which we are looking at and so once you do this you can put in the numbers and you can extract now this probability of given a certain level of this cripple gene where am I in the embryo so here's my cripple expression level and here I can read of where am I in the embryo and I show you for 3 examples of how this is done so basically here when you have this level of cripple and so now it means that you're confining your X to this little region if you're here your probability distribution is too peaked and you're confining your X in either of these two so now it's ambiguous you're either on the upward or on the downward slope you don't really know just by measuring this one gene and if you're measuring down here well you're measuring a lot of noise and you barely have no idea where you are and so this works well for one gene but as soon as I want to add now 2, 3 or 4 genes I can't use this notation anymore because I would need more dimensions in order to visualize it yes no actually from 4 to 1 all I want to know is where am I in X along this axis here that's good I'm happy for the challenge what will I not always find I may not but let me maybe finish and you'll tell me in the end if I am so with this one gene I am not with this one gene I can barely with more or less precision I can tell you if you're here I can more or less tell you if you're here I have more precision on each of the slopes but now it's ambiguous I can be on the upward or on the downward one so with one gene I cannot okay but now I want to add more genes in order to add more genes I need a different way to visualize things we are using a so-called decoding map to do this and that is it's the same it's just a different representation of the same kind of data all I'm doing here is I'm basically I take these probability distributions and collapse them and put them on the Y axis and so if I'm basically asking well if I measure a certain amount of X what is my probability that I can find a certain position and so I basically put myself here and let's say I want to know what's going on at 0.55 well there is a probability here in the center and so that maps me somewhere in that region here okay it's a so-called decoding map and sorry as examples you have same three examples here if you're reading yourself off here well you're somewhere in that region in the center if you're in the upward or downward slopes you're either on this side or on that side and if you're measuring out here in the noise you barely have any idea where you are okay but this method of representing things I can now expand to many genes and here's essentially the math behind this so we're making the approximation of the two embryo fluctuations in these gene expression levels they're Gaussian and we have reasons to believe that that's a very good approximation because we have done earlier work of where we actually we can actually show that and what that means is that your probability solution of finding of having a certain this is what we measure right of having a certain set of genes given a particular position a a what do you say an expanded Gaussian to four dimensions so if you only would have one gene all you would do is to put here g minus g minus the expression level of the gene minus the mean of the gene you square it and you divide it by the variance but if you expand this to four dimension that variance becomes your covariance matrix and you multiply again by by by the expression level minus its mean and because you have four and it's a four by four matrix you do this on both sides with two different vectors so this is just an expansion into four dimension of your regular Gaussian and then once you have this you construct your lookup table for any embryo alpha that is your sample now that you want to see where is it in your dictionary you take that embryo and basically find the most likely position X that's consistent with the state at which the embryo was given its expression levels okay so it's not a model it's just a pure probabilistic way of estimating position and all the probabilities come from data and the Gaussian approximation but we have very good reasons to believe that the Gaussian approximation is a very good approximation at the level at which we are looking so this is not a model this is purely based on data and probability driven okay so now expanding this to two genes going back to the data so now we are adding to this yellow guy which was peaked in the middle we are adding this guy which was roughly on off and now you are look up to your your map, your dictionary look such that you are still ambiguous and imprecise in certain regions but you see that you are populating more and more levels here on the diagonal which means that your map becomes more and more precise and by the time you go to four genes you see that along this entire length of your embryo now you can determine this very good precision of where you should be given those four levels of of genes these four expression levels of the genes and if you ask how precise can you do this, well that precision is of course the width of this probability distribution here and you can look at this width along the entire length of the embryo and what you see is that you recover very nicely the sigma x which was constructed from a completely different in a completely different approach so here, but what we gain of course is that here we can decode meaning here I can now give you, or you can give me four concentrations of the gap genes and I tell you unequivocally with 1% precision where you are and that 1% precision we have reason to believe that's mostly due to what the residual biological fluctuations are okay, any questions so far? yes so a single cell is kind of your limit of what you need okay and so doing any better than this would be kind of redundant in the system you could of course add more genes here what we show in this case is that it's not needed that with four genes at least between 20 and 80% egg length you're good to go okay and you could also have more cells and in that case you probably would need more genes, yes I agree yeah no, exactly it's a good point there's no new information all we have now is that with this new approach we can also decode but the information that I gave you with the old approach and with this new approach so far is exactly the same I'll show you in the next few slides of how we use this now to test the system and to really see whether it cares about this 1% you had a question? exactly, yes so you mean you I'm not sure I understand so you're saying if I'm looking at 1% lines in between those 1% lines I get precision for free this is a local function and I compute it at each x yes, exactly I use only local information no it is used in terms of the covariance matrix no sorry you're right what is not used is correlations in fluctuations across space we have not used those at all all we have used is correlations in fluctuations locally how they influence each other we have not used spatial correlations that is true what does this mean in the correlations of well maybe maybe not so we have to check so we saw a long range correlations also in space and we think that the system is tuned to a critical point but that's for a different talk but we can talk about it later we see some correlations also in space all right so let me test the system now and so how can I do the sorry were there other questions here before I continue so how can you test the system how long do I have because there are so many questions already during the talk I get those good I'll be fast so what you can do is now you can ask so how do you test the system a way to test the system is to present the system with a different distribution of gap genes and see if based on the different distribution of gap genes it still determines the output the same way it did in the wild type case in the wild type case I tell you give me four concentrations at one position and I tell you whether there is an e-stripe or not or whether you are on the edge of that e-stripe or not okay so if this is really working and if the biology really cares about this 1% precision then I should look in an embryo where the distribution of gap genes is completely different reconstruct this entire coding book and then use the same concentration those concentration of 4g go into my coding my decoding book or my dictionary that I constructed from the wild type look in there where do these four concentrations are and I should say there is an e-stripe or there isn't and so the flyer allows me to do that because what I can do I have these three different maternal inputs and I can take them out one by one I can take all three, I can take out two I can take out six different sets of two and I can take out one at a time so I have nine different conditions that I can look at it's no time I'm just looking at at one given time point a four minute window no time okay so and then I can go in and take a wild type measurement of these different stripes it looks like this here so in blue you see one, you see the mean and in light blue you see different embryos and that tells me where in my code book is the combination of genes that has a stripe and so what that means the first thing it means of course is I need to reconstruct the dictionary in my... sorry I need to reconstruct the distributions in the mutant backgrounds and for this you basically have these different conditions in fact it's only seven so in this case you take out one at a time in this case you take out two at a time in this case you take out all three so there should be very little information in this system and here and here you see again the wild type the movie that you have already seen okay so these are the GAM gene expression levels in mutant backgrounds and you see they're very different and so these differences we can now use to test our system and so the first thing let's see if this one here really has no no information and lo and behold it's a flat line and the actual output in the mutant is also a flat line there's not a single eave pattern and the mutants if you take out all three gradients that have maternal information that have information about position okay which shows you also that it's that those are the ones that you need because if you take them out you have nothing left now let's look at an example where we just take out one okay I take out one of those maternal inputs and you see that the differences in this one in this case I take out one that's called torso it doesn't really matter what it is but you see that the distributions are slightly different they're subtly different but they're slightly different for instance this red bump is completely gone this bump here has shifted to the left these guys have shifted to the right so there are subtle differences and now to look at the decoding map of this guy you see that there are lots of regions where it is following the diagonal but towards the edges it deviates and if I now take my wild type measurement I would predict to have no stripe 7 here and these 6 stripes I would predict at these locations and if you look you see that indeed you have stripes at these locations and what's really remarkable here is that you it's a very subtle effect the deviation here is maybe a percent or so but we get that effect correctly now I give you an example for a gross effect so if you take out sorry this works for one of those 7 stripe genes and it works for all of them now I give you an example for a very gross difference so you see that the distribution of gap genes is very different from wild type if you take out this bicoid mutant first of all they all are seemingly halved but then also their relative concentrations are very different to each other and now the decoding map looks like a mess there's nothing on the diagonal there's a little bit of schmutz around here and there's a little bit of schmutz however if you predict again you're dead on where your mutant patterns lie and quite surprisingly you also detect this guy here which is a famous duplication of e-thripe 7 because this guy here this pattern here is a duplication of that one so not only do we get very subtle in fact that are like 1% effects but we also get very gross changes here we predicted a change that happened at 50% different position you are very unhappy back there tell me you don't want to tell me why you're unhappy with my data? oh you're very happy no you were the one telling me that's impossible so but just to summarize this works for I showed you two examples in detail this works for all of them and now you can summarize prediction versus actual location and you see that they all line up very nicely on a diagonal and the deviations are again at the other of 1% or so so I guess I stopped here I showed you all this but I'm not going to conclude and you're not going to see anything about tomorrow I'll show you live stuff