 Good morning everybody. Nice to see you again on the second day of our summer school. Today the team will be a little bit in trying to obtain measurements which we can use then later on to fit to our models and the relation between getting information and models while yesterday we were more talking about the pathophysiology, so we go one step further and in this regard it's my true pleasure to introduce James Sharp. He's Nikea, a research professor here at the Center for Regulatory Genomics which is connected to the UPF, to the university here, and he's leading a research group on Organogenesis, and the nice thing is that he saw just as several of us are seeing is that the most important is that we integrate knowledge as what's going on with modeling, with trying to get the right amount of image data in order to try to understand some biological processes, in a way in a systems biology way, and that's of course where modeling is really really important but at least as important I think is getting the right data and even in order to get the right data sometimes you have to develop your own imaging techniques or improve imaging techniques and that's exactly what James is doing, has been doing so we're looking forward to what you have to tell us. Thanks very much. Can you hear me? Yes, okay. That seems to be working. Great, so what I'm going to be talking about today is specifically Organogenesis as just mentioned, so how an organ grows during embryo development, and so I think it's quite having a look through the schedule of this week I think it's a bit slightly different focus from what quite a lot of the week is about because it's not explicitly about tissue engineering or adult tissue or well differentiated tissue but how in the first place you build an organized structure and as mentioned as well, yes I'm coordinating a systems biology unit, so this is one version of what we think of as systems biology, in particular the integration between data and modeling. So what I'm actually going to do is try to summarize essentially the kind of journey that we've been on over a while now actually, like nearly 15 years I'm not going to go into lots of details but I'm going to cover quite a lot since this is I guess a school and the more you see them the more useful it is to you so what I suggest is just interrupt me at any point if something's not so clear because I'm going to go through lots of different things so in a way it's the journey from starting from a very crude attempt to build a model of limb development up to what we have now and a lot of it is to do with building tools, so imaging tools on the one hand but then also modeling tools, so computational tools so what's interesting for us about multicellular systems, I mean systems biology is already trying to cope with the complexity of single cells and as you know they are sort of horribly horrendously impossibly complicated, for at least these kind of reasons the network of interacting molecules, whether proteins for correlations, gene regulation, whatever, have very complicated circuit diagrams, topologies per se a lot of the interactions are non-linear and there's tons of feedbacks and those three things alone make it very difficult to understand or build a model of a single cell and yet what we want to do already is to make it even more complicated and try to understand multicellular systems where you have another level of feedback so because cells are communicating with each other we have not only the feedback of circuits within a given cell but the feedback between cells a cell signals to its neighbour, can change its state but its changed state can signal back to the first cell and all kinds of interesting and exciting and complicated things can happen as a result so immediately we get into spatially extended situations from ODE's to PDE's and essentially it's a lot of extra feedbacks so what do a multicellular bunch of cells have to do in development? well as you know essentially they all come from one cell at the beginning and so early on you've got lots of cells but they're all the same so not just that they have the same genome, so the same wiring diagram, the same potential but they're even actually in the same state so the parameters of the system and the variables of the system are kind of uniform and what these cells have to do is to somehow specialise into different cell types but in a spatially controlled way so a kind of paradigm example of this would be to develop a stripe of expression, let's say the green gene in one place so this might be the head end of an organism, this might be where the legs are going to be and this might be the tail or this might be different regions of a limb or different regions of an organ how do you get these cells when initially they're all the same to make a coordinated choice that will be different from the others and that question alone, just how you design a circuit that's in every single little cell that will do this kind of thing is a complicated question in its own right so we have been doing quite a lot of that and here's like a little model, computer simulation of a bunch of cells which are responding, now in this case they're responding to a morphogen gradient so there's something that's already asymmetrical which is something that is very a big issue in developmental biology where do the asymmetries come from and which sort of asymmetries lead to which other asymmetries but here let's assume that there's already something different at one end there's a smooth gradient and then the network, the program within each cell is to somehow interpret this gradient so that's quite complicated but of course during normal embryogenesis or organogenesis it's more complicated than that even because while these cells are communicating with each other and making decisions they're also moving, they're also changing position with respect to each other during this whole process now that is often seen as a kind of hierarchical or sort of one directional system, information flow we tend to think of the genes at least you know molecular biologists, genes and molecules as the clever bit the sort of brains of the system or the electronic circuit of the system every single cell has one, they're making decisions, they're communicating, they're deciding what to do they can tell at sort of the next scale up cells what to do to migrate in certain directions to divide in certain directions and if you have thousands or hundreds of thousands of cells all being correctly controlled individually then together you'll get a global result like tissue movements and shape changes to build something like your arm but of course it's easy to see that it's not a one-way system at all and if in this toy example if we had the cells at one end and the other end were already secreting some repressor of the green gene but a diffusable repressor that's going to flow across the field then actually just the geometric rearrangement of the tissue could allow that the green gene gets switched on in the middle because these cells have become further away from these ones than they used to be and the actual point of growth is that the relations of the positions of cells change so in fact there is a feedback from tissue movements, this kind of macroscopic scale down to molecular regulation and it is in fact a feedback at a larger scale so we've got at least three levels of feedback in a developmental system in a system that's starting from ball of cells and trying to build organs many many complicated feedbacks within the cell we've got feedbacks between the cells and this would be true even if nothing moved but because everything is moving and this movement is controlled by the system itself we have a higher level of feedback and that's why we decided that to understand a particular example of development, the developing limb we would try to tackle all three levels which honestly is not being done so much in developmental biology and it's why we've focused a lot on new imaging technologies for this kind of scale and ways of deriving or inferring what the movements are so we're interested in all of these but we've spent quite a lot at this level and quite a lot at this level not as much at the single cellular level actually but I will show you a bit of that as well so this is our example structure that we're trying to understand the vertebra limb or mammalian limb and in fact this starts just as a little ball of cells on the side of the embryo in humans this is about a four week old embryo and in mice it's 11 day old embryo so in fact even in 11 days it's gone from a single cell to a very complicated structure but the limb bud at this point is more or less just a ball of cells maybe about 50,000 cells at this point and within it these cells have not yet decided whether they're going to become bone, cartilage, the fingers, the tendons, whatever and during just a two day period they make tons of decisions so again bear in mind there are thousands and thousands of cells around here this is the expression of Sox9 which probably many of you are familiar with and this is a labeling of which cells are expressing Sox9 so the critical thing is that a cell just here has decided not to express Sox9 and the cell here has decided to express it and here it has and here it's decided not etc so since all these cells are very close to each other how do they know to make the right decision? this is of course prefiguring the entire skeletal arrangement of a limb so this is a critical process so essentially we can think of this as two types of question the molecular patterning how do cells know which genes to switch on and off through all this signaling and interactions and also how does the correct physical shape of this structure emerge since it starts off as something we almost know asymmetries just a sort of ball of cells and then it elongates in a certain direction then the hand plate widens so that you'll get fingers developing there so how do these two things happen? so we can go back to our scheme of multicellular systems in general and in our case we look at this all with respect to the limb so molecular patterning we can look at gene expression patterns in limb buds and there are potentially hundreds of genes that can be relevant to this process cellular activities and cellular behaviors and the whole tissue movements and in the lab we have projects both data gathering projects and modeling projects at all of these stages and what you'll also see when I go through some of these examples is that we don't yet have a model even a sort of beginning of a model that yet does this whole cycle so we believe that cells know how to behave because of which genes or molecules are being expressed or in their environment conversely the tissue movements are due to cell behaviors so this is sort of due to this this is sort of the causality is due to this and these movements also affects dramatically the gene expression patterns so far the examples I'm going to show you are more like are mostly a single arrow like if you know the cell the 3D or 4D distribution of cell activities can you predict the tissue movements or if you know the full description of tissue movements can you predict the gene expression patterns etc putting it all together is what we're still sort of just in the process of trying to do and again in terms of technology for getting data in some ways what we're looking at is these types of almost like omics or going towards omics trying to get lots and lots of gene expression patterns but again the important thing is this is all spatial and so imaging is the key I mean essentially imaging is just I mean taking a picture is just getting a 2D or 3D distribution of certain variables over space so whereas in developmental biology again images have mostly been used just to illustrate the point where it's really about quantitative data and then the other thing is that it's not just a case as some projects seem to be of gathering as much data as possible putting it all into a computer kind of cranking the handle and seeing that something will come out one of the reasons why we chose the limb bud is because it has a long history like decades of history of being studied as an example of organogenesis which go back way before the molecular biology revolution and in fact this really helps because it means that the questions the conceptual questions to be addressed are very clear which is not so much the case I feel for many other organ systems that have only been studied more intensively more recently for example how does the elongation occur I mean it starts as a kind of a ball and it elongates very specifically in one direction it's not growing in all directions how is that actually controlled at the cellular level and also at the molecular level how do you get what we call proximal distal patterning or specification how do the cells in the hand know that they should make hand rather than your mid arm or the upper arm because essentially these cells are all the same and in fact what we do know is if you transfer cells the early bud cells from here to here they can make this structure they're not predefined that these cells have to make a hand they're somehow sensing their environment and making decisions with respect to their neighbors so this is proximal distal patterning and the other one that we've worked on quite a bit which is quite fun is digits how do the cells make this periodic pattern of being sort of specifying towards making a digit and then no digit and digit and no digit and digit etc this is digital patterning and again there have been debates going on for years and years and even decades about how this process works and then in the end of course the value of maybe building a computer model is that we can understand these processes not just as individual concepts but how they all interact with each other because they're all happening at the same time growing specifying along this axis patterning along this axis they're all simultaneous processes and in fact what's particularly confusing about that and I think we'll not be understood until we have a good model is as I said we've got these different kinds of processes elongation, regionalization, digit patterning and other things, shape control, scaling but if the organ was 20% smaller everything gets 20% smaller if the embryo was 20% bigger everything gets 20% bigger so there are different features that we can observe work in a coordinated way well it's complicated partly because it's multi scale as I was saying but also complicated because the main signaling pathways that we know of that also many of you I'm sure are familiar with like FGF, Wnt, BMP from all kinds of functional experiments it looks like FGF has roles to play on all of these processes so in one bunch of cells there are different FGF patterns and it's somehow influencing all these different processes the same for Wnt and the same for BMPs so how are we really going to disentangle the cause and effect, the control of this system and the idea is sort of by putting it all together because if we have different processes like morphogen interpretation versus periodic patterning versus cellular growth this is what we think is involved in this process this is some of the molecules and part of the circuit involved in this process which I'll be explaining and again FGFs and Wnts involved in cell motility and polarity but you'll see that basically many of these molecules are the same molecules so FGF for example and hox genes and even Wnt they're involved in all of these processes simultaneously so there seems to be a kind of very complicated circuit that simultaneously does many different things or that we at least conceptualise as different functions the other thing that's sort of quite interesting to know as a background to this whole thing of modelling development modelling multi-cellular systems is that it actually has a very long history especially for limb development so when I went up to Edinburgh it was explicitly to do a post-doc to start trying to build a computer model of limb development and the first thing I discovered is that in Edinburgh where I had just arrived before I was even born just someone had already done it and even published a nature paper and this was in the 60s so people were realising that this was the way to go decades ago I mean we kind of think that with all our clever computing and stuff we're very modern and of course we are but it was clear already a long time ago that this is what had to be done and some brave people were trying it so this is the computer that this simulation was done this is not the real simulation this is just a picture the simulation looked like this this whole computer had 24K around the output was on this printer here which was these little asterisks printed and they wrote something and it is just remarkable for this paper back in 1969 that they had decided to explore whether morphogenesis can be simulated in some of its important aspects using a digital computer because in those days a computer was still sometimes a person where the cells are represented by numbers stored in the computer and their genetic instructions are represented by parts of the program I mean it's prophetic because that's what we're still trying to do now they also put very sort of modestly because of the limitations imposed by the methods of printout and core store capacity this 24K as I discovered there are very compelling reasons for restricting the model at least in the first instance to the study of growth patterns in two dimensions only and we're still publishing two-dimensional models actually decades later the other thing that I have to admit and it's really important the conclusions that they came to in this paper have turned out 40 years later to be completely correct there were a bunch of models of this process in the intervening decades that are wrong and these guys were completely right so it's not about the idea of building models or even the ability to build simple models theoretical biology is nothing new it's about integrating data so I think what has happened is not just that computer power has increased and whatever but the fact that we have all these amazing ways of getting accurate and quantitative data so I'm going to show this whole scheme now with sort of these two halves gene expression, so the kind of molecular side and then the physical or geometrical mechanical side tissue movements and they're kind of linked in both directions in this direction how gene networks make patterns over space and time is very influenced by how the tissue moves so that's the link in this direction but of course the link in this direction is maybe more obvious and more studied which is that molecular changes molecular processes both the gene expression but expression of morphogen gradients expression of all kinds of molecular states is what tells cells what to do actively what kind of movements to do and then that together creates tissue movements so what we really want to do ideally is get data which is going to be imaging integrate data and then actually do the modeling and I'm going to go through sort of present it in this kind of a layout and I'm going to talk quite a bit about imaging tools that we've developed but also about computational tools so the first thing was how to capture the 3D geometry of an embryo like this and it might by now hopefully seem like a strange question because now we've been able to do it for quite a while but literally back in 2000 when I was first trying to do this there was not a good way to do this I was taking these limb buds and cut them into smaller pieces already because a whole limb bud is too big to image under a confocal microscope and then try and reconstruct these pieces in a computer and this is not going to be very satisfactory for a model if you really want to predict things so I had to try something different something new and it kind of made me realize that at that point there was what we would call this imaging gap it was very easy to image very small things like microscopy, microscopic things cells and tissues and of course large microscopic things like human patients were also quite reasonable to be imaged but it was really not clear how to image something like this they're too big for microscopes and they're too small for MRI or well for CT you can do it in various ways but then you might lose molecular information so I decided to try the idea of X-ray CT but using light and as some of you will know the way that CT works is that you shine some kind of rays like X-rays in this case through your sample you capture a quantitative shadow the shadow from one orientation does not tell you where anything came from so you can see there's a hole here the light a bit and there's a darker bit but from one view you can't tell where it came from this is sort of the opposite of most microscopy techniques which focus very sharply on one plane and eliminate noise from above and below but with tomographic reconstructions you just do this process over and over again at different angles and you can then calculate using this back projection algorithm what the distribution was in 3D so the idea was to try this using light instead of X-rays and luckily for me because I'm not a technologist or a physicist or an engineer sort of officially but nobody had tried this before so this is essentially what I tried putting an embryo supporting it in agarose clearing it and then just taking pictures of it at different angles and then pretending that these pictures which are really images they're sort of diffraction limited focused images but pretending that they were shadows and seeing whether this would reconstruct using a standard back projection algorithm which would be here so this one plane here all the data from that plane will go onto one row of pixels one angle looks like this you can't tell where anything's from and as you reconstruct it as you accumulate data from different directions you can simply calculate what's on the inside of the embryo and I always keep showing this because to me this was just magic I mean this raydon transform and this back projection but this actually worked and of course because you haven't cut the embryo you just do this for every row of pixels and you get full sections slicing all the way through your sample so this was fantastic I mean in the department where I were a team of five people had been physically cutting embryos and trying to reconstruct them for a decade and luckily this made things much much easier not only that but you could image molecular distributions so this is the kind of thing that you can do and here we have just fluorescent antibody labelling against neurofilament and again in terms of size you know this is a 12 and a half day mouse embryo it's probably about 8 millimeters from here to here and it just hadn't been possible before so this was very useful for what I wanted to do in fact it went way beyond what I really wanted it to do because I just wanted to image simple things like limb buds but in fact it's becoming a tool for mapping gene expression patterns over many different kinds of samples and species here you've got two different genes at three different time points and we can't really call this spatial transcriptomics but we can map and integrate data from many different genes and then map them in 3D so this was one of the exciting kind of outcomes of trying this idea and it meant that as you've seen we could map both expression patterns so something that's molecular and shapes over time but what about getting dynamic data so that was all kind of fixed samples so we did do a project to try adapting this OPT Optical Projection Tomography to work on living limb buds and you can see an example here because essentially it worked but I'll explain the caveat in a moment so here's this process that we're interested in the gradual development of the skeleton in the limb bud here is a mouse that has GFP under the control of a gene that's expressed in this way that represents the development of the digits here we cultured it in the machine and from that rotation you can get a 3D image and then luckily within the machine it actually managed to develop so you can see that here by the end 19 hours later the fingers have developed I mean which means they're not fingers of course it just means that some genes have been switched on in the cells to form the fingers this is just about patterning not about differentiation and that meant that we could indeed get a 3D movie over time of this process so again this is digits 4 as it happens 4, 3, 2 your fingers always develop in mammals in this order that 3 first then 2 sorry 4 then 3 then 2 and this extension of green is not cell movement or tissue movement it's genes switching on a gene sorry cells switching on a gene so the cells are just sitting there and then they're just deciding to switch on this gene and the same here and the same here it's not about flowing or movement of tissues or cells so in principle we could get all the data that we want like this but this is just not practical for a mouse developmental system one thing I mean this depends on making a transgenic mouse that is expressing GFP under the control of a gene and we can't just make lines for every single gene we're interested in well and there's another problem that I'll come back to it in a second but the other sort of limitation of this is we don't here have single cell resolution so we're not tracking individual cells because remember there are about by this point three or four hundred thousand cells and the resolution is not high enough to track individual cells so if we see shape changes which we do how much does it tell us about the underlying tissue movements well in fact in this kind of cartoon you can see if we know that this shape moves to this shape if that's the only thing we know we just know the surface it doesn't tell us much about the movements you need some kind of landmarks because there are many different ways of this shape could warp or twist or morph into another shape so we need some kind of landmarks and we did that with live OPT as well which is just on the surface putting fluorescent beads and doing 3D reconstructions of those and then over time you can see the limb butt growing in the machine as well and you can see the movements of these little beads and since we have them all in 3D we can track that but this is just the surface doing this for the internal tissue there are methods that you can label randomly stochastically cells internally and we have managed to do that to a certain extent but it's proven very difficult and the main reason is this the main limitation for imaging mouse organogenesis at really organ stages so these are like embryos of E 11.5 or 12.5 they don't live or grow very well in vitro very young mouse stages can but at these stages they just don't we originally were doing it like this taking a piece of the embryo and skewering it and then sort of orienting it with this device and in fact some processes continue but real growth doesn't continue very well so we did then explore a bit doing a whole conceptus method where instead of taking a part of the embryo you take the whole embryo plus the yolk sac plus the placenta and keep everything intact because actually the thing missing from here for this limb bud is blood flow there is already by this point a beating heart and a circulatory system and without blood flowing through these limb buds they will just not grow and I think this of course is a design tissue engineering that somehow getting blood flow capillaries and flow through tissue in vitro is going to be essential to get above a certain size so we tried this and in fact this works but it is really tricky to do this so this is what I just described the whole embryo you can see the heart is still beating you might even be able to just sort of make out in these capillaries here of course you have to have the whole thing you have to have the placenta intact as well because the circulatory system is going through the placenta as well and it worked because we got really good limb development in vitro that way but it is just too much of a pain so we essentially had to go to a whole different idea or philosophy for a lot of our data capture which is what I will be talking about now which is how to infer dynamic things from static data it is kind of frustrating if you were working in Zebraficial Drosophila you wouldn't dream of doing this because you can image those embryos alive but mouse embryos don't grow normally outside of the uterus so so we did do some of this and this has contributed to some of our data using this whole concept as culture with live open tea and we can get shapes and dynamics but really we are going to have to focus on static snapshots like 3D snapshots of all kinds of data so then the next kind of problem became apparent and in fact of course this whole project, this whole talk is essentially like a series of problems and how we overcame them this is something that I am not sure how appreciated it is embryos if you harvest them even from one litter from one mouse are not the same age and if you are trying to understand the process it is very dynamic hour by hour assuming that all the embryos that came from one litter from one pregnant female assuming they are the same age is going to be completely wrong these are all from one litter and this stage is very much behind this stage in fact with the system that we have developed these are almost 24 hours difference in developmental time and just coming from one litter from one mouse so one of the first computational tools we had to develop was a staging system and we chose to use the changing shape of the limb bud over time as the metric for development without going to the details we had to harvest and take photos, 2D photos of hundreds of limb buds and try to work out what is the average trajectory of this shape change over time so here you can see time going from the bottom up to the top from the point when there is no limb bud on the embryo to gradually a time at the end where you have got a clear outline of a limb bud with the five fingers formed there and in fact this is a publicly accessible website if anyone actually wants to stage embryos it's I think for the stages that we've got it's the most accurate way of doing it statistically we can show that you can stage an embryo to plus or minus one hour by just taking a photo of the limb bud drawing a spline around it and this will tell you the age it gives you a result kind of like this with a confidence limit and then you can actually go and analyze in real detail the temporal evolution of all your gene expression patterns sort of hour by hour so that was one of our first tools there that was necessary for dealing with static data now how are we going to actually take how are we going to get a continuous description of these tissue movements from static data and for that we've switched into 2D some of our projects are in 3D some of the projects in 2D and we kind of jump back and forth depending on what's necessary or convenient we've got shapes at many different time points as I've shown you and we realize then another piece of information that we could use to define how the tissue moves is clonal data clonal data or clones means labeling a single cell at a certain time point and then seeing where all the descendants are maybe 24 hours or 48 hours later and then integrating that somehow so we've developed this thing that we call a morpher movie because it's a movie in the sense that it's just descriptive, it's not a model but it's descriptive of the changing shape and movements over time and we started with this which I've already shown you and from this we can make simple little models like this so this is just a kind of a simple 2D mesh that's remeshed every hour and if we have a hypothetical distribution of movements you can label a single point at a certain time point and see what's the probability distribution of a space of where the descendants of that cell could end up if that map was correct so the map that I've just shown you is just invented hypothetical but if that map was correct then a cell that started here the descendants would have to be roughly in this distribution so from that we were then able to create many different maps sort of hypothetical maps of how the tissue moves and compare it to real clonal data so we digitized all of these clones we have these maps all as digital things and then you basically compare the map with every clone and look for the best fit and the best fit allowed us to produce what we think was the first accurate description of the in 2D of these movements over time for the mouse limb body and it all came from dead fixed data static data so that would be how to describe this side tissue movements a continuous description of these movements how are we going to get a continuous description of the genes switching on and off and their expression domains and the dynamics of that well essentially using a similar thing so we can of course get a gene expression pattern at any moment but we just need some way to integrate these with the movements that we believe are happening and I'll show you how this works so it's easy to do what's called a whole mount in situ hybridization for any gene you're interested in this is actually the benefit of looking at gene expression patterns rather than proteins with antibodies if you want to look at where a protein is distributed you need an antibody against it and for most proteins there are no functioning antibodies for a gene it's much much easier you just know the sequence and you can make a probe and you can do it in situ so any gene in the genome can be analyzed in this way and at different time points and it's expressed in these cells and not expressed in these cells in fact it's not just an on off thing you have kind of gradients so we then developed another little piece of software that for any given time point of our map we can take a gene expression photo and align it for that particular time point and kind of map these levels these are not qualitatively accurate they're kind of qualitatively suitable let's say I mean you can distinguish between high expression medium expression and no expression so that's one time point and one gene and in fact because of the automatic staging system we can make this process a bit more automated now but what we really want of course is many many time points and many different genes so here are three genes A11 and MACE mapped over time these three genes are of interest because these are ones that are believed in early development well in this developmental phase to distinguish between your hand and then the forearm and the upper arm HoxA13 basically specifies the cells that will make your hand A11 this part of your arm and then MACE this part of your arm so you can make little movies of that now I emphasize that these movies these are not models yet this is just a representation of data because later I'll show you movies that hopefully look very much like these but they're very different because it's actually a dynamical predictive model so that's data capture again all from images and we now have been able to integrate data on tissue movements and data on gene expression patterns can we do all of this in 3D we're kind of doing this all in 3D now but it not surprisingly takes a lot longer in terms of the movements we're essentially using a similar approach that I described a little bit we're using the live data that I mentioned from live OPT but in particular we're using again lots of fixed clones but 3D clones in this case are much more controlled ones so going briefly through that you can see here of different limb buds of different ages with clones of different sizes that means that a particular cell was labeled at different time points in the past and all the descendants are labeled later and to image that we've been using another imaging technique that I'll just show you because it's related in terms of applications to OPT in a different way so if we've got these kind of rough scales up here from hundreds of microns to millimeters to centimeters OPT what I showed you is great in this kind of size range but another technique called SPIM or light sheet microscopy is actually good for this smaller end of the mesoscopic range and we've been developing this technology as well along with a number of other labs so this is an example of the kind of thing that SPIM can do, light sheet microscopy watching dynamically a process in a zebrafish embryo it can also do fantastic static reconstructions of slightly older mouse embryo stages so this is a mouse head again labeled for all the nerves and then OPT though is still more suitable for things like this which is a project looking at diabetes and mouse models of diabetes where we were able to analyze many many pancreases in an efficient way to quantify beta cell mass and stuff like that so this is just an overview of the kind of techniques that we think of as mesoscopic imaging now in between microscopy and macroscopic imaging so with SPIM we were able to pinpoint cell by cell in these clones and you get something like this this is one particular clone here we've got every cell it is as I said before it's unfortunate that we can't do this in a live sample but not only because it's difficult to grow embryos but also when they're alive they're not transparent enough so from here instead we can map clones from many different experiments and each clone of course as in the 2D case has got some information about how the tissue must have distorted to make this grow and to cut the long story short there we're kind of using various kind of inverse methods and finite element modeling to fit what must be the tissue movements to satisfy all of this data over time and space so that's really I think all I'm going to say about all this capturing capturing data and even about integrating because I've shown you these kind of tools that we've really had to work on for years and years to integrate all these different types of data so we've got the imaging done and we've got the integration done so what about actually asking with all of this some scientific questions and I'll just go through briefly a couple of our modeling projects to give you a feeling of how we use all of this well the first one elongation there was a very longstanding theory called the proliferation gradient hypothesis this is a limb bud and it's known that at the tip of the limb bud certain growth factors are secreted and diffused so you get a kind of a gradient with more of the growth factor at the tip and less further back so the very kind of simple idea that had been around for about I think about 40 years by the time we published this paper was that these growth factors simply stimulate proliferation they just stimulate growth but nothing more is really specified than that so this would be a kind of isotropic behavior a non-uniform but nevertheless isotropic so more growth here in a unit time well so what we did was just to test that using these kind of data and images we used finite element modeling this particular package and we took two kinds of quantitative data one is the actual measured 3D distribution of growth rates which you can do with this technique a sort of pulse chase experiment with BRDU and IDDU so we built 3D maps of the cycle time and then the other data was of course the real shapes from OPT imaging but when you put these two things together and try and run the model just as a forward kind of problem this is the limb bud at an early stage this is the real limb bud shape 6 hours later and when you run the model it doesn't do anything like the reality it basically blows up like a balloon it doesn't go in one direction now we did lots of things for this paper that I'm happy to talk about later if anyone's interested so we did it as an inverse problem as well to optimize what pattern of cell proliferation could theoretically explain the shape change but essentially the conclusion from everything is just controlling growth rates per se isotropically cannot explain limb bud development this is another kind of comparisons and the kind of ironic thing here and I think it's a very important thing to worry about is that there were already in between the early days there were three papers published on modeling limb growth which all they didn't prove or anything they just basically adopted this idea because it was in the literature which was that growth was concentrated at the tip concentrated growth would be sufficient to explain the elongation and they made models of it and the models work and they made many different types of model this one was a finite element model this was a kind of vertex model and there was a cellular POTS model as well all published and all saying yes this is how the limb grows but this is not how the limb grows it's easy to make a model that will just replicate what you think is happening but of course it doesn't prove that it's happening it's maybe even counterproductive because to the then experimentalist biologists who then read these modeling papers they think oh well my ideas even being shown in a computational model so that's even more support and this is just not true why did they all make I mean how did they all make a mistake because they didn't bother to actually fit it with real quantitative data it was just more a conceptual model and conceptually the idea that it's possible that you have more growth at the tip and that you elongate but when you compare it with the real data it's just not possible so it made us realize that there must be something oriented going on some anisotropic cell behaviors or activities so this was our first example of a kind of modeling project here and very briefly it forced us to go back to do more kinds of imaging and develop more kinds of things that I won't go through to somehow watch these cells in a developing limb bud and as I mentioned in the mice it's just not doable so we switched to chicken embryos because they're more or less the same to watch how the cells move during limb development and came to the conclusion where I'll just realize by looking down the microscope they're not randomly oriented at all they're very strongly oriented in certain directions they're very complicated cells in all kinds of funny ways and so this led to our next hypothesis our next hypothesis was that and is currently still that essentially there's a convergent extension process going on that for any little block of tissue cells are trying to intercalate all the time this would lead to an expansion in the other two dimensions but because of the arrangement of the blocks and I'm going through this quite fast but just to give you a flavor of it because of the arrangement of the tissue and the arrangement of this intercalation it allows that the limb bud doesn't expand in these directions but only elongates in one direction so we then went back to modeling to explore that and in this case we have actually used a cellular POTS model which is a type of model that can represent fairly arbitrary shaped cells it can be done in 3D and in this case the people that actually were doing this modeling, Julio Belmonti in James Glazer's lab they had been recently developing an extension of this process whereby cells could grab onto each other at a certain distance and pull in order to represent a kind of intercalation process so for example here if this cell 5 and cell 2 pull towards each other they will pull they will push cells 4 and 3 apart so you will get convergence this way and extension this way you then have to specify how these links will be orientes with respect to gradients and in this case it's a Win 3 gradient coming from the ectoderm and in this way we could test our hypothesis and essentially we can just show which doesn't prove that it's right but we can show that this hypothesis at least is consistent with limb development so on the side of tissue movements and mechanics and things this was another modeling project that we've done so here with finite elements here with cellular POTS modeling and in fact in the lab we're not at all we're trying not to be wedded to any particular modeling framework or formalism we just try whichever ones are useful regionalization how you get how the cells know that they're in the hand versus here versus here this we've been doing gene networks we've been doing so far mostly in 2D because it is more complicated I showed you already this this is simply mapping how these genes behave over time but in the literature it's fairly clear that these genes seem to be controlled by upstream gradients and upstream gradients of retinoic acid which is coming from the body and another upstream gradient of FGF which is coming from the tip so you've got 2 opposing gradients in opposite directions and the question is if these are the inputs and these are the known outputs what is the structure of the gene regulatory circuit which will do that conversion that will take these inputs and make these outputs and for that we've been developing kind of reverse engineering approach so we take a hypothetical network we've got our real output results here and of course this network requires parameters so we just take random parameters and simulate and these three genes these three colors are these three genes and you can see that if you just take random parameter values you get patterns that do not fit very well the reality but because this data is all digitized and because the model is of course a digital model we can do automated comparison between the output of the model and these gene expression patterns and again to emphasize this is comparisons over space and time so we're comparing at every time point during development for every gene and for every triangle how well they fit and then you can put that into a reverse engineering kind of parameter optimization loop all through iterations the result gradually improves such that here we've got the best result for this circuit and we then explore many different circuits and see which circuit is the one that would fit best and this is the result the circuit that seems to fit best and this is the actual dynamics so the movies that were showing you before was just digitizing the data but this movie is a digital PDE model where we just give the initial conditions the parameter values and it grows like that not just the initial conditions I have to confess the growth movements here are not being controlled by the genes just to be clear the growth movements are given as well so then I'm just going to show you the final example of a modeling project again on this side of the pattern and it's quite a fun one because it's about a debate that really was going on for ages and ages in the field which is how you specify your fingers and really it's been a debate between the principles of local self-organization versus positional information which were both championed in a way by these two different people this one much more famous than this one in developmental biology Lewis Wolpert is very famous but Alan Turing is of course famous to most people in the world although most people in the world don't realize that two years before he committed suicide he actually published a paper on morphogenesis and his interest was actually in developmental biology it wasn't some kind of mathematical curiosity actually all of a sudden obsessed with how embryos develop and it always seems from in the hindsight of these days where we have developmental biology and computing and things as distinct fields weird as to why he would be interested in these two different things I mean one thing you have to remember is that computer science didn't exist because I mean he was one of the ones who invented the computer that wasn't really a distinct thing as you probably know as well like the Turing test artificial intelligence he spent a lot of time trying to work out what intelligence meant that's how he developed the idea of an algorithm which was a formillism of solving a problem but in later life he started to realize as they had various attempts at building computers what exists in the universe as the most clear example of a very intelligent machine and it's the brain and then he all of a sudden at the end got obsessed with trying to understand how does the brain build itself how does a brain as a physical object come into existence in the world and for that reason he got interested in development and for that reason he published this paper which was absolutely astounding kind of breakthrough it started a whole little field of theoretical pattern formation in biology and what did he discover he sort of discovered stroke invented that if you have just a system of two molecules that diffuse but also react with each other in a single cell this doesn't look like it could do much and it probably can't do that much it could be homeostatic or it could oscillate but it can't do that much you've got an activator and a repressor but when you have this distributed over space so you have this same system every single point in space and you allow these to diffuse in just the right way this quite remarkable thing can happen in this little toy simulation the colors represent the concentrations of one of these and you just spontaneously get a periodic pattern so the key thing here is that the reason why these cells switched on the gene and the reason why these cells switched on the gene is not different it's just a local self-organizing patterning system you've got in a way three stripes here but it's not because they're in different positions the cells don't really know anything they're just reacting with each other the dominant alternative view in the field has been that in developmental systems there are such things as morphogen gradients or other kind of things distributed over space of the cells and cells read out this very little information and therefore decide what to do so in that kind of framework the cells here would have switched on SOX-9 let's call it because they know they're at position one and the cells here would have switched on the same gene but because they know they're at a different position and the cells in between would not have switched on the gene because they know that they're at the position in between and where did this idea came from it came from very very convincing experiments the famous ones in development back in 1968 where from a chick wing if you take a little piece of the very early wing so days earlier than this when you've just got a little bud if you take a piece from the posterior side and graft it onto the anterior side you go from a normal wing which has got these digits to a full mirror image duplication and that looks to all the world like this kind of French flag model where at different positions from something different distances you specify different things and because you've now taken part of this organiser and put it on the other side you've got a gradient this way and a gradient this way and therefore you've got more fingers and you've got all the different identities now this cannot be totally wrong but I'm not going to go into that right now the point is that these were two very very different explanations to explain a periodic pattern of digits now we've been doing many things but I'm just going to show a couple and this is a mouse limb bud in green is where Sox9 is being expressed and one of the kind of observations that's been around for ages but we've just done it in a more kind of modern way is this that if you take cells even cells that have not chosen the gaps between the fingers but then you culture them in micro mass and I'm sure that some of you are doing similar kinds of experiments they will spontaneously switch on Sox9 but not all the cells but not in a random pattern but actually in something that looks a bit like a Turing pattern I mean if you don't take just the negative cells if you take cells mesenchymal cells from the limb bud in general Turing patterns but even if you take the cells that have apparently chosen not to become cartilage not to become fingers they will switch on Sox9 if the fingery neighbors have gone maybe even more remarkable if you take just the Sox9 positive cells and culture them some of these cells will switch off Sox9 but again in a patterned way not just in some stochastic random process not very great, I've got to get a better movie of this but it really looks to all the world like a local self-organizing system also highlighting that Sox9 I would say is not a sort of a differentiation marker it's a gene that's involved in the dynamic process of deciding which cells will become fingers and which will not and it's very dynamic and it can be switched off just as easily as it can be switched on so we then use these kind of things to screen which are the molecules the big question that all the biologists want they don't really care that much about the model they just want to know what are the genes what are the signaling pathways so we use this system to do differential microarrays and RNA-seq and transcriptomics and things to find which genes seem to have patterns that could be relevant for this stripey process and in particular which signaling pathways because the Turing system depends on having at least two diffusible molecules and the diffusible molecules in these cases are signaling molecules our conclusions are that essentially it's BMP pathway and the WINT pathway I won't go through the details now but I'm happy to go through them if anyone's interested but we have very precise time course of how different aspects of these pathways get switched on and as a control for example FGF does not, does not have any stripey pattern and from that we then had to construct a computer model and again I'm not going to go into the details but we've essentially been playing around with different ways of screening all the possible simplest mathematical models that can fit the data in terms of the phases of which node in the model whether it's BMP or WINT is in phase or out of phase of the other genes and construct a model the result of all of that is this which is one of the criticisms that people have always given against Turing systems for explaining fingers is that a Turing pattern is not a controlled thing I mean it just goes all over the place and if you run it ten times you'll get ten different results now this is completely true if we run our model with no extra control in our limb bud growth model you just get a labyrinthine pattern which is a typical Turing pattern furthermore if I run this ten times you'll get ten different patterns but to cut the long story short what we gradually realise through many kinds of things and many kinds of experiments is that genes that are controlled in space and time have an influence on this Turing network so for example Hawks A13 or D13 which is expressed in the same place they have a specific pattern over space and time which is represented here if we let that pattern actually influence the parameters of the Turing system you go from this to something like this in this case where it's restricted so that the Turing pattern only occurs within the Hawks D13 expression zone this is still not very reliable and it's not what fingers look like but if we take another one which is the FGF gradients and allow them to influence the Turing pattern we can get a totally reliable model where just the five stripes corresponding to your five fingers pop out they actually pop out in the right order with digit 4 coming first then 3 then 2 and then 5 and 1 later you can see a comparison here so this is our model and you can see we're not caring about the long bones at the moment we're only caring about the fingers but if you look at the time course of Socks 9 it matches very very well so in fact our conclusion about this debate which has been a big debate for a long time is that to pattern something like this you need both it definitely depends on this local self-organizing process that is sort of similar to or maybe equal to a Turing system but a Turing system alone is absolutely not enough you need some more sort of spatial control on the system which you could call basically positional information so that's our last example that I'm going to show you so what I've basically shown you then is you know how we've integrated lots of data how we've integrated it how we've been developing imaging tools but also computational tools but then most importantly is the kind of more fun bit is why to do all of this because we spent years doing all of this but the reason is to address actual clear conceptual questions that are unclear in the literature the other point is that of course this is not really a one-way flow like this putting data and then integrating and modeling or whatever I mean this is just one useful way of looking at things but I mean science doesn't work like this at all what's really missing here is the fact that you make predictions and those predictions tell you what experiments to do next to try and break your model and you go round and round like this so this is also what we think of as a key thing for systems biology and then the final thing is to integrate this model but you know and it potentially seems great because it's linking a molecular feature to a macroscopic phenotypic result why would you believe it well based on what I just said about experiments and feedback and whatever just gathering tons of data and putting it into the computer is not enough you do have to try to break the system and test it but for this case we have managed to do that in various ways can this model where we just took a lot of data made a kind of model, simulated it and got some results can it really explain perturbations and of course what you would want to do is to take each node of your system and perturb it, perturb it in experiments perturb it in the model and see how well it can predict things that were not going into the model in the first place and I'm going to show you that this is why we do believe the model to a large extent, not because of what I showed you but what I'm showing you on this slide in a series of hox mutations, which is essentially like an elix series because there's four hox genes that are expressed in this region as you knock out more copies of the hox gene in mice these are the phenotypes that you get in terms of fingers you can't see them very well here but you can get mice with 13 fingers and they're almost completely normal apart from that and it's going very linearly as a sort of a correlation with knocking out more hox genes the more hox genes you knock out the more the phenotype goes in this direction until when you knock out the final one you actually don't get any fingers and in our model if we do the same and gradually reduce this we get the same result the FGF for our model to work the FGF actually has to have an influence on the wavelength and it seems a bit abstract and maybe speculative to say that well FGF would have an influence on the wavelength of this Turing system that's modeled quite abstractly but in fact when we've done micro mass culture and put increasing amount of FGF into the culture we do get increasing wavelength in the culture which is exactly what would happen in the model finally down here the more sort of interesting results is that we can culture limb buds not enough for them to grow well as I said but enough to show the patterns of the fingers developing and the model predicts that if we inhibit the wind pathway instead of getting individual Sox9 domains which for individual fingers the fingers start to grow and merge and grow and they sort of fuse into one big finger region and it's actually exactly what happens in the experiments if we inhibit BMP we should get the opposite results that the fingers disappear and they do and that BMP expression gets up-regulated which it does but probably the best is that the reviewers of the paper then forced us to do something more because you could say this is just up-regulation of a gene it's not really about a spatial pattern it could just be boosting something or knocking it down and we had to agree that they're right because the whole critical thing about spatial patterning and Turing patterns is the spatial dimension so what we then were able to do actually is by having combinations of these drugs at different concentrations we were able to show that both the model and in reality you predict that you changed the wavelength of this pattern and I'm not showing it all here but the beginning of this experiment where you've just got two big fat digits is where you've already got four fingers starting you can just see it in the GFP faint pattern and then you put these combination of drugs reducing BMP signaling reducing wind signaling and the pattern reorganizes so certain cells that were off switch on certain cells that were on switch off other cells that were on stay on other cells that were off stay off I mean it's nothing about movements it's that the cells start making different choices and they rearrange the pattern within the field of cells so this is why we actually believe that the model must have something to do with reality despite being you know a very simple abstract model well and is it relevant to congenital abnormalities probably yes because for example hox genes are well known to be causes of these kind of abnormalities in humans so lastly is that all that gene regulation stuff that I was showing you was in 2D but of course we're working on a 3D model and this is just to show you how it's going as it were I think there's another one over here this is the expression pattern of hox A11 over space and time and again this is not the data this is actually the dynamical PDE model over time so as I stressed at the beginning I mean what we've achieved so far is studying these different concepts separately but you could even argue that some of those concepts could be explored with simpler models the real goal though is to get them all into the same space into the same model because for sure they interact with each other digit patterning affects PDE patterning and vice versa they affect growth growth affects the patterning so the idea is as I was showing at the beginning to get all of these processes going simultaneously in the same model so that we can understand the complicated kind of interactions of cause and effect between these things this is this is the kind of PDE model that I showed you before this is the Turing-like model although this diagram is from before we knew that it was BMP and Wint and this is the growth that I was showing you which again is controlled in our model essentially by FGF gradients and Wint gradients and now we think that Wint is one of these molecules so all of these things really are sort of tangled up together working simultaneously so I would just like to thank my lab for doing all of this work because obviously they did all of it and thank you for your attention thank you very much for this fascinating story let me first start with a couple of more what I would say like in brackets non-scientific questions in the sense like what you show is like since you have more like originally a biology background and you've come to all of this how do you decide or how do you organize it to go to these models do you at some point money, hire engineers and physicists or do you say you're going to talk to different groups how do you see this interaction between biologists, physicists engineers in order to get all this working okay so yeah that's a good question I mean the first thing to say is that I was more originally into kind of computing than biology but as a kid really I mean I was a total nerd and into electronics and stuff like that so in fact for me the beginning was sort of understanding how difficult it is to get systems that do controlled behavior and getting systems to do what you want but then I got interested in biology because of course you saw kind of the genetic code and everything being quite like a code and etc so then my formal training was all in biology my degree and my PhD but and it was totally experimental with no computational stuff but I was getting frustrated all the time so I did five years of making transgenic mice and doing experiments and the whole way through I was just frustrated that there's no way we're going to be able to understand even the experiments that I'm doing without a computer model so I decided just to make a big switch at that point and to go specifically to just myself to do a postdoc in modeling which funnily enough one of the postdocs in my PhD lab told me it was a scientific suicide I mean literally that's what he said because biologists or experimentalists really did not have much faith in modeling and maybe for good reasons because I tried to point out that a number of models have been published with absolute disregard for the data and are pretty unhelpful so anyway I made that decision and then I was looking around at places to do the postdoc and talking to people and in the US as well and the sort of general feedback coming back and I felt this anyway myself is you've got to do it in one lab collaborating I mean there may be different experiences here but back then which isn't like nearly 20 years ago people was confirming what I was suspecting which is that you know I have to do both I have to do the modeling and the experiments so I decided from the beginning that I will well that was my postdoc but then when I became independent to build a lab that would have you know would be really interdisciplinary and and it sort of worked pretty much from there but as a result I have always just employed directly myself into the lab, business engineers and computer scientists and you know I mean at the moment I have I think six scientists and two computer scientists and only three biologists and that kind of works I mean the problem of collaborating as many of you may know is that what is a useful experiment or set of data for a model is not necessarily a kind of experiment that will excite an experimentalist on their own so having the two together and really packaging them as a sort of single interdisciplinary project to me seems to be the best way so the direct communication between the engineers the experimentalists or in case of for example medicine the medical doctors is essential in order to get these things I think that is quite an important message also for most of the people so really go to the data is very important now partially related to that and you also pointed it out is like as you see as you say like there have been models published that in the end I would say turn out to be rubbish it is like in my way of speaking but partially it also it determines the field like in the beginning you say like you choose limb development because there was a lot done but on the other hand there is also to me at least there would be huge bias in for example reviewers at both sides that said like okay this problem is solved namely I publish this so whatever you are going to do you can't get published and that is an experience that some of our people also already have is like if you try to do it differently because you say like look whatever you published actually doesn't fit the data so you should do it differently every reviewer is going to say like no no no I mean this is solved you have to use this type of approach because that's what everybody is doing I don't know if you have similar experiences and how to deal with this like when be stubborn and go to this kind of data change the world in a way or when just kind of give up follow whatever has been done I don't know what I would give as advice I just I mean our own experience is clear and maybe a bit similar to what you are suggesting which is that at least so I started being interested in modeling these Turing patterns more than 10 years ago we didn't publish it until relatively recently but at the time when we started the field was absolutely dead against that and you know it wasn't formalized because we were not publishing anything and there was nothing discussed formally about this but it was clear that people really hated it so but I mean what can you do if you believe that something works in a certain way you just have to follow that so I knew that it was well I knew that there was a chance that it would be difficult and some of the other things that we've been trying to publish and push at the moment have also been essentially against received wisdom and it does make life more difficult but I you know I think you just have to go with what you believe in so I think that's another very good message because a lot of us are trying to do things nobody believes in it reviewers say like no don't do it if you believe in it I think it's important just keep on doing it any other questions okay thank you very much very nice presentation I have one question when you in the experiments of grout I would like to know if you have observed a variability in your measurements or more or less the measurements are under the same conditions you obtain very similar grout response I would like to know about the variability of the experiments and if this variability exists have you thought how to incorporate this in the models well we have not really addressed variability I have to admit I mean in a way we've you know we've started with even an assumption that there is such a thing as the average normal growth trajectory and so then we've taken whatever data we could and averaged it into one growth trajectory how far that's true or not well remains to be seen with more work but I mean it is true that one advantage of studying a process like a developing organ is that I think it's reasonable to say that you know a developing mouse limb is a developing mouse limb and the things that we want to understand in the first instance are just how a mouse limb develops per se in general because even that is far from clear I would only really go to looking at variability after we've got at least some basic mechanistic understanding of the average behavior but it is true that for example compared to things like studying tumor growth I mean there's a massive advantage here because no two tumors are exactly the same so you would not be able to build up an atlas of normal tumor growth because it's just not a controlled process whereas you can do that for something like organogenesis other questions ok then I want to thank you again then it's coffee time now so thanks for being with us