 So I would like to welcome all to this virtual computational biology seminar. Today we have the pleasure to have Jörg Stelling leading the Computational Systems Biology of the Department of Biosystem Science and Engineering at the ETH Zurich, but in Basel. Jörg studied biotechnology at the University, at the Technical University of Braunweig in Germany with an intermediary stay at the Köln-Normal-Sperre d'Aconomie in Montpellier. He received his PhD in systems biology in 2004 from the Department of Mechanical Engineering, University of Stuttgart in Germany as well. And his PhD thesis devised new methods for the analysis of robustness in complex biological networks. And in 2005 he joined the ETH Zurich as an assistant professor for bioinformatics in the Computer Science Department. And since 2008, Jörg is an associate professor for computational systems biology in the Department of Biosystem Science and Engineering, as I say in the ETH Zurich in Basel. And his group is also part of the SIB, the Swiss Institute of Bioinformatics. So the Jörg group comprises biologists, computer scientists, engineering and mathematicians who perform interdisciplinary research in systems and synthetics biology. The group focuses on developing and applying computational methods and mechanistic mathematical models to study complex cellular networks. The group, biological application, rely on the group experimental biology part that uses yeast as a, budding yeast as a model organism and on various external collaboration. So today Jörg will tell us about multi-scale models for the design of synthetics gene circuits. Jörg, thanks again for accepting this invitation and the floor is yours. Okay, thanks again for the invitation and for the nice introduction. I picked a topic that I think I didn't talk about yet here in Lausanne, which is different from the normal work on metabolic networks or other things we work on. And so the idea here is to use mathematical models in order to construct new biology. So this synthetic biology type of approach and I'll show a few examples on synthetic gene circuits today. So the basic idea of this field, synthetic biology is essentially to industrialize fabrication in biology, design and fabrication in biology. And so that eventually you will do most of the design using computer-added, aided design methods, using large libraries of parts you can use for the design. And then eventually you send this design, test it virtually, send it to a robot, it gets assembled and you make very complex products. So that's the vision or the theory. And really the engineering ideas here are shown like this. So the idea is to operate with modules, so well characterized parts that have defined interfaces, defined input-output relations. And then you can do plug-and-play as you would do with software for example. Okay, so that's the idea. And also many of the analogies for this field are drawn from engineering, a lot from digital circuit design. And so think of this here as a hierarchy, you have a computer chip at the lower level, you have transistors, but you don't want to design a computer chip based on the individual transistor characteristics. And so you have intermediary layers of obstruction like logic gates and from this you construct chip and from this you construct software based on top of this. And so the analogy with biology that is often made is that, okay, so the signaling pathway for example would be based on molecular interactions, but the signaling pathways would act like logical gates but only in biology. And the open question is essentially at which level we have to use which type of description and in the end whether we can really do this design in a modular way as we could do in software or traditional engineering. So what you can think about in terms of how to engineer organisms now, if you start with a wild type here on the left hand side, simple things that are also done routinely in biology for example are just modifying individual interactions, for instance knocking this one out and making a new connection in this existing network and you can think about integrating a new part or trying to make these things a little bit separate, partially autonomous and the other extreme it would be much more synthetic so the ultimate thing would be to build a synthetic cell that is completely reprogrammed. So this would be the wild type, this is the most synthetic circuit. So the problem with this approach is essentially that in biology these modular ideas, nice interfaces between different parts they actually do not really hold. So if you have say a circuit like this here so this symbolizes genes and then essentially what happens is that they are not autonomous when the minimal interactions with the cell or the host are required for instance load on ribosomes on polymerases on energy. So there is some sort of context dependence unlike in electrical engineering, at least in digital electrical engineering. Another question you might ask is if you want to revive existing cell signaling circuits and you want to simplify a cell, so how can we do this? Can we do minimal replacements of complicated biologies for example here? And the third question pertains to this picture here. So this is the host cell, what happens beyond a cell? So if you have multicellular organisms you want to do say synthetic biology approaches for curing diseases in humans then you're definitely not acting at the cell. So these are the three questions I want to address with sort of examples today. First of all to what extent can we design something that is autonomous, well controlled entities in a biological system. Next question, okay can we understand essentially design principles or what would be minimal functions we would need to implement and test this. And the third question is can we scale this up to essentially higher levels, control entire organisms in the end. And please interrupt at any point if you have questions. So there's a microphone here as well. Okay so the first part and this is the caveat here for the entire talk. So we use a lot of mathematical models, these are wrong by definition. And what I want to show you is that even rather simple models in this box quote here can be useful for our purpose on that design. It's not understanding in detail what biology does. Okay so the first part, specific examples what we try to do is look at this example of whether simple biological components, in this case a transcription factor that we try to design, whether this can have predictable behavior even if it has to interact with a host. If it needs column raises and ribosomes et cetera in order to execute the functions and of course it depends on the host environment such as the host growth rate. So the basic idea here was the following. So we want to have yeast cell where we have a synthetic design transcription factor, this one here, that we can influence with an extra signal, in this case better astrodial, such that we can control gene expression of a target gene in a very precise and reliable manner. So the design of this transcription factor here is essentially three parts. It has a DNA binding domain, lex-a, that comes from outside yeast. Binding to the promoter has an activation of this estrogen receptor domain to interact with astrodial, so that's the input processing domain and it has an activation domain that binds to the RNA polymerase. So the idea is if we add astrodial it will get a predictable output in terms of gene expression for this target gene. So that's the nice picture that you put as figure one into publications. The reality is rather different if you look at the interactions between these different parts. So we have our transcription factor that is supposed to control the open reading frame, the expression here. This requires polymerase, polymerase requires polymerase to make polymerase. This all influences is affected by growth. So this nice linear pathway here actually doesn't exist in reality. And the question now is can we try to capture the main interactions here with the cellular infrastructure and still make this behavior predictable? So what we did was essentially construct these transcription factors, several variants and use mathematical models and I'll show you the basic structure in a minute to evaluate these different design variants. They all differ only in this activation domain. So the DNA and RNA polymerase binding domain, these two transcription factors here for example and the points here are the experimental outputs as a function of the input so this beta astrodial and time and we measure fluorescence reporter. And you see even if it's exchanged just this domain here, behavior is quite different in terms of magnitude but the surface here also tells you that the model quite accurately describes this behavior. All together we build four variants and the other two are shown here and interesting part of these two here is they have much stronger activation domains and this eventually when you activate these transcription factors too much over time will lead to cell death indicated by fluorescence that goes down over time. So there's an interaction definitely even if you over express a single gene that can drive cells to death. On the other hand you may change conditions in this case just changing the medium and this will change growth rate for example in glucose growth, cells grow faster than glycerol. The effect is essentially a change in this expression capacity where you can see that the model captures this quite accurately. So what is behind this model is essentially a relatively large reaction network so it's quite detailed as an ordinary differential equation model based on mass action kinetics. So for example if you just pick one of these reactions here so this will be the transcription factor bound to its target then polymerase comes and they interact with each other. So 30 states, 30 variables, the interesting or the critical part here is to explicitly model this infrastructure like polymerase and then what we did we used part of the experimental data in order to identify the associated parameters 45 and then test it against independent data and the data I showed you before these are relatively large data sets so we had roughly 450 data points for training and then twice as much for testing and that's actually what I showed you in the previous slides. So then the question you can ask if you just use the model and use it for predictions for example you can ask can you predict the growth rate as a function of the input and for the four different transcription factors and this is shown here so those are the two toxic ones and actually you can predict this quite accurately so there's really a interaction between what happens at the single gene level and the cell and then the model offers you the possibility to look at what happens in more detail so for example that's the total polymerase concentration here you see that for these toxic transcription factors this really goes down simply because the transcription factor captures more polymerase there's less for making new polymerase essentially and those response characteristics are not necessarily trivial and then last up we asked what can we do in order to tune essentially this system here so you might change the affinity to the operator that's what is usually done in engineering transcription factors but you see the effect, the relative fluorescence so you change by a factor of two or so at most whereas when you just multiply the number of these binding sites here then you can actually achieve something that is pretty much linear so you can really ask how you can modify this quantitatively okay, as I said if you have any questions just ask yes so we see variability and also what I showed you here is this FACTS data here I show you just the mean and variance so you see over the population that's not huge in the next example we come to that no you cannot do this in general in this particular example it works and the problem is a little bit yes in principle yes you can do it but then if you want to do it in vitro then you really have to have the right conditions that are the ones that apply in vivo otherwise you measure something that only holds in vitro this is one example so this has exactly the same parameters as the model if you do this say in more glucose, metabolic state compared to a glycerol, metabolic state you may simply estimate your parameters well so let's go to a bit more noise so the question, so this was just a single component now we're trying to go towards a circuit and the question behind with this was can we do minimal replicas of natural circuits I'll put it the other way around if you can make a replica that has essentially the same function as the natural circuit this gives you an idea of what is really important in the natural circuit so we have two examples they're both related to metabolic control and the basic feature of those is we have uptake of a nutrient, so this is E. coli, this is yeast, cerevisine and we have uptake of a nutrient that is normally not present so the uptake systems are suppressed when for example glucose is there but if there's enough in this case lactose or galactose the control system switches this on then the cells are either in one of two states either no expression of this entire system or fully on this works by having a transporter that is expressed expression is controlled and essentially here we have a number of intermediary steps that process the signaling and we have feedbacks so the basic structure here in E. coli is simpler than this one here in yeast and the main question we asked, okay so what's essential here is this part essential, is any of these here essential or which of the feedbacks are essential so when you overlay this and try to come up with a common design here then it's essentially this one, we have something that induces the uptake we have the transporter, the permease this then will lead to this inducer here in the middle that relieves the repressor that represses permease otherwise there's no positive feedback overall in this circuit okay so that's what we tried to replicate just this little positive feedback circuit and the way how we did this, we built this all synthetically so with components, so this is again in yeast with components that do not exist in yeast and the first step was making this replica here so essentially the circuit diagram without the feedback this is all constructed here, we have a reporter's item and we just characterize the steady state input-output so the dose-response relationship for this particular constant okay so without the transporter which is this one here you don't get any fluorescence, no uptake of IPTG in this particular example and otherwise you get the fluorescence here okay so from this again we try to identify a mathematical model for this part and also later use this to tune the circuit and then we just took the same model and asked what happens when you close this feedback okay so the circuit is now a little bit extended the only thing that is different here is that instead of the reporter we have the transporter here and the reporter is essentially in parallel so that we can see what happens in the cells so the previous configuration so this linear one was open loop then we have the closed loop with this feedback here and what is shown here in these pictures are essentially as a function of the input concentration these are cell population distributions measured by facts fluorescence and this one is the experimental data and this is the model simulation okay so we use open loop data in order to identify the model then put in both in reality and in theory this loop the feedback and then we ask what would happen and what you see here in this experimental data as well as in the model this is actually get signatures of a bistable system a system that can exist in two different states so you have a subpopulation here that has low fluorescence another one has high fluorescence and eventually this moves to full fluorescence another hallmark of this switching device is the following you can start with high concentrations of the input or start with low concentrations and the system should have dynamic memory it's called hysteresis so for instance here at this concentration here if you start with low input then the cells will stay in low input if you start with high input then some cells will move to low output but a lot of them stay in the high state so there's really memory and hysteresis in this system so that's yeah so in the experimental data for the closed loop you have that interesting bimodal distribution starting for 0 microgram and 1000 microgram in this window here do you have any suspicion where that comes from or is it some other interest? it comes from this system here when you close the loop that essentially for the same input it can exist in one of two states so either high or low at the individual cell level this depends on the history and now if you have a mixture then some of the cells may stay in the high state if they started in the high state some others may go to the low state and the reason for this is that we have stochastic gene expression and so what happens here in the experimental system when you for instance make an upshift you start with low concentration and then go to a high concentration then cells will definitely react differently in terms of when at that extent they induce expression that's the corresponding simulations from the model and this also is different from whether you start high or you start low so what we can do then now is for example this type of experiments that's in silica now you start from a stationary one of these bimodal distributions split it and one part here is off one part is on and then you ask what will they develop over time what you see here is they converge to the same stationary distribution again and then you can ask what are the switching rates so the rates of random transitions between one and the other state and you cannot directly infer this from the experimental data so this is shown here so essentially these rates depend on the input concentration and you can infer this using the mathematical model so in this case definitely stochastic gene expression is the most important thing okay so the bottom line of this here is what this tells us is essentially in these complicated real world circuits this small loop here may be sufficient to explain at least this basic behavior of having different phenotypes of having memory in the system and on the other hand it's a new design for making a switching knowledge clear or questions okay I have a third part and this now really goes towards this question how do we scale from a single gene to a circuit to essentially an entire organism if you eventually want to control what the entire organism does so we modify something in a subset of cells or in a single cell and want to influence the entire organism so that's something that in engineering would call embedded controls so you embed a controller into some device like a washing machine or whatever and here we look at this in terms of medical applications so we've been working on this for quite some time so this is now in mammalian cells so this is collaborative work with Mark and Fusnega and just to give you a little bit idea so what we try to do here is essentially just building a time delay device in mammalian cells but that's really just at the cell levels you want the cell to switch on to a particular output say the production of a particular target protein only after a certain time delay or building an oscillator again in single cells but nothing interfacing between different types of cells or the cell and an organism and that was essentially the first thing where we thought about this we have essentially two types of cells these ones here so they are essentially sender or receiver cells they do both of this send signals, metabolic signals in this case we have another type of cell that processes the signal and gives the signal back so this is now going from single cells or single cell types to engineering essentially something that may in the end look like an ecosystem if you do this for microbes or something that is relevant for medical applications so the critical part here is really how to either avoid interfaces or the natural system and these two left examples or to actually engineer it in a way such that again you get predictable behavior predictable control so the example I want to show you in more detail is related to type 1 diabetes treatment so probably everybody knows that there is a chronic disease and that it has huge health costs so the underlying reason here is that essentially there is no production of insulin so in there better cells are gone that produce insulin so the normal physiological control is that insulin essentially leads to uptake of glucose and glucose metabolism when there is too much and there is interactions between the glucose and blood and brain so better cells sense glucose secrete insulin depending on the glucose level and by this achieve a steady state of glucose in the blood that's the basic control okay and especially so for type 1 diabetes patients they have excessive blood glucose because they cannot produce insulin and this also has potentially drastic consequences because then cells that need energy may switch to fat-based metabolism instead of glucose-based metabolism if the insulin dosage is not right so what you get is essentially accumulation of acids and this is called acidosis this can be potentially lethal especially if this goes over there goes to the brain so it's really important to control this in a tight range the blood glucose levels on the other hand what you can do is actually you can try to explore this mechanism here in order to sense what happens in the blood system and the reason why this is important is the following so treatment options for type 1 diabetes are essentially insulin injections either manually or with automated systems but they still have a lot of problems one being non-compliance or that the system simply do not really work or that they work they adjust dosage just too late so what you might want to design in terms of biology is essentially a cell that acts like a better cell so senses glucose and produces insulin appropriately and that it can then encapsulate so mammalian cells encapsulate them in some gel and then implant them under the skin and they do the control really in real time in a physiological way that's the vision the problem with this is simply that there are no in order to do this you have to sense glucose in the blood but there are no appropriate glucose sensors now so in the first iteration of trying to design something that works at least in this control loop works like a better cell we looked at the following idea it relies on what happens in metabolism when there's too much glucose so said then other cells in the body will change their metabolism switch to fatty acid metabolism essentially and this will lead to an increased acid load in the blood stream so there is a natural buffer system so that's bicarbonate in the blood but this only holds for a certain range so eventually there will be a lower pH so too many H plus ions in the blood and there may be ways of detecting those so you detect the effect of insulin misregulation indirectly by measuring the blood pH that's the basic idea and this is now the resulting circuit here so this is a essentially a pH sensor so this is engineered comes from a different type of cells this is then wired into a natural signaling pathway this one here so protein kinase A signaling in mammalian cells and then we take the output of the signaling pathway and reroute this to control a genome interest so these mammalian cells are modified only by expressing the sensor here and by expressing a sensor for this signal here and then producing a protein of interest in response and then the problem here and the reason why this requires really quantitative description here is you have to have exactly the right response in the right pH range to produce eventually the right amount of interest so then we we try to do this using a lot of mathematically modeling and that's the basic idea, the basic setup so we have these modified mammalian cells that I showed you in the previous slide and they have this sensor components engineered and they control gene expression but they also have again their own metabolism et cetera so what you can then do in vitro just with these cells is for instance modify the medium pH and see how they respond try to capture this quantitatively and then design the cells in such a way that they would do the right thing also when they are in the body so we use this part in vivo characterization with the reporter protein here in order to build a model and then try to use this model exactly the same model to predict what happens when you now produce insulin instead of this one here and for that we have essentially a very simplified model of what happens in the body so it's a concatenation of different models from physiology this is now from the mouse with the appropriate interfaces but in the end what it does models the response to insulin how this affects glucose then glucose will affect the body metabolism in terms of how much CO2 is produced which is important for the buffer and how many acids then we have a model for how this will affect the pH in the body and this feeds back to the tracentamic cells okay so that's really multiple scales going from these single cells to the entire body and the challenge is to have a model that is good enough such that you can do the identification here and then just plug it into another model to predict what happens in vivo okay so just to walk you a little bit through the steps here so again the lines are model simulations the dots are data so first of all the idea was to characterize how the sensor is responsive to pH changes there's red bar here that's the natural physiological pH and essentially here you see that it does what it should essentially it increases protein production when the pH is a bit below the natural pH in the butt and this is an experiment and simulations that tell you that this is actually reversible so here you change the pH it goes up and you change it again it goes down essentially so this is the type of characterization experiments we have a signal transduction pathway in the middle we have to a natural signal transduction pathway so we have to characterize this one as well so what is done here is essentially two different pHs the sensor should support differently and then you get increased protein production in the one case so you have the full dynamics and it's low in the other case and this is now essentially a dynamic experiment where we vary the dose by modifying the induction time and when you put these two experiments together what you will find you cannot just describe this natural signaling pathway just as a linear type of signal propagation what you need to do is describe it in this type of way so we have an input here X which activates Z so for instance this protein here but there are some dynamics here in the middle that essentially leads to something called an inquiry and feed forward loop so what this does is you release the input here eventually the system will adapt always to the same system so that's now inference on what happens in biology but we need this part in order to describe the data just to give you an idea of what type of experiments you can do characterizing in vitro so this relies essentially on the interplay between acids and CO2 which is the buffer in the blood you can also modify this in vitro change calm dioxide concentration to add different buffer concentrations here and the point about this slide is mostly that in many cases the model characterizes these is very much consistent with the data so then what we did this was used to establish this model here and then we plugged it into a mouse model so a model representing basic physiology of the mouse and there were parallel experiments and experiments so this is now essentially the sensor characterization in terms of an output this is insulin now but when the cells are implanted in the mouse and you see that with the same model the behavior is nearly the same and the data and model again look pretty consistent the implants were then done in healthy mice and type 1 diabetes mice the circuit is called pH guard and what happens in healthy mice essentially you have this pH guard it changes the insulin a little bit up but not much different from the natural when you give placebo but in type 1 diabetes mice the placebo essentially produces no insulin whereas this one here produces insulin and essentially here you see the time courses time course data so this is for the natural response to glucose injection the glucose level in the blood so that's how healthy cells would respond that's how type 1 diabetes cells respond so they start from very much higher concentrations and they essentially have a hard time down regulating this again and this is now the circuit here so that's the experimental data so the mice with the circuit and the model simulations as you see the related dynamics also has very much changed and then do just use the model ask how safe would it be in real implant situation if you for instance don't know that there is zero production of insulin in the native system so would this affect insulin levels or glucose levels would they still be in the tolerable range and the answer from these many simulations is essentially that yes you would not predict that for instance these insulin levels reach extremely critical so that the pH goes extremely down and just two more slides on essentially second generation of this type of ideas the idea is now to use direct sensing of glucose instead of the surrogate signal pH so the way how real beta cells do this is the following so when there is low glucose then there is essentially no metabolism of glucose and metabolism via ATP interacts with calcium and potassium channels in the cells so when glucose comes ATP is produced this leads to inhibition of this calcium-potassium efflux the cells depolarize this will lead to influx of calcium this will then lead to release of insulin and natural beta cells and the idea here is that we should just take different cell types and engineer them in such a way that we can follow the same logic and the interesting part here was essentially that only this this calcium pump needs to be included and a particular transcription factor in order to recapitulate the beta cell biology and you have to do it in the right amounts to separate quantitatively so just the result essentially the main result is just in this configuration so we only have these two constructs here synthetic constructs we challenge this with exosolid glucose then insulin is produced and this is now time course data the scale is days over all the experiment was 6 weeks if you have wild type cells this is again glucose concentration in the blood they have really normal concentrations if you have the control cells for type 1 diabetes you see this huge high insulin levels in the control mice and the reason why the time course here is so short that this is simply most of the mice die under these conditions and when we use this implant which is called ag1 beta the blue lines you see essentially you make an initial perturbation when you implant this to the mice but eventually it goes to reasonable levels of blood sugar and the mice live for at least 6 weeks and the final curve here so this is implant with beta cells derived from humans and you see essentially that's there a little bit worse than these engineered cells okay, that's essentially it so what I wanted to show you is that you can actually do model based design even if you we don't know a lot about the biology the question of how much biology you have to include or what level of abstraction is possible really depends on the context or the purpose of these models and in terms of methods what I showed you especially in the last part this is sort of an art how to do this type of model development and interfacing all this the question is how can we do this more systematically that's really an open question and then what I wanted to show you is this idea of having physiological feedback implants with these designed cells this may be an opportunity for applications but also in terms of theory because essentially these models I showed you all that they try to describe what happens in the entire mouse they're relatively simple because simply the organ has a lot of control circuits itself they may get away with very inaccurate models okay so the people who did this work from my group are listed here so this, especially in the last part was a longstanding collaboration with Martin Friesenegger's group in this NCCR molecular systems engineering and David Auslein and Mingyi Shi were the main people on the experimental side for the mammalian work and that's the final statement