 So I'll start out and I'll end with the people who are actually doing much of the hard work that is Harsha Surbith and Nisha They they put in the work in doing the coding and doing a huge amount of data curation and analysis as you will see and Then there are various members of the consortium who I will also mention later Okay, so let's consider some kind of pathway model This is a block diagram. This is not individual molecules These are each of those has got lots and lots of individual molecules in it so we actually had to work to come up with a way of in fact zooming in and looking at it I Said to them come up with a Google Earth like thing so you can zoom in and see that see how horrible it really is But do so in a manner where you don't lose track of the various pieces So you can zoom in on any one of these if you so desire and each one of them is extremely complicated So that was just about visualization. So this is how we did it some 20 years ago This was a much smaller model. It looks awful because we didn't have the the ways to zoom in on different parts of it But the problem is that for a long time People have been kind of stuck That is a size of model with maybe a hundred or so reactants and reactions Seems to be in where people have plateaued in their efforts to come up with this kind of detailed signaling reaction Kinetics-based model. So why have you been stuck all this time? So it's not because of computer power computers can do And even then could do pretty much much larger calculations It's not been because simulators can't do it the it's for the machine is just a matter of scaling it up It's not been because we lack the data or the capabilities to get the numbers What I think the the problems have been are the the following few Which is first of all these are all these models have been handcrafted. We've gone out individually dug up the various citations worked hard to fit them into the model and That takes a lot of time and effort Having done that if you want to use it in a different context or you want to mix it with something else That turns out to be another whole exercise in hand crafting and rederivation and that's not easy not been easy When you Builds build a model you would like to then say okay I can reuse it wherever but the trouble is that signaling pathways the chemical reactions involved in them are Sufficiently tied up so it's sufficiently interconnected that you cannot just take one model and plug it into another one Which would have been a nice clean engineering solution. It is harder than that and Finally and perhaps not the least It's extremely boring and extremely costly to go out and get the numbers Whether this means hiring a bunch of people to go out and mind the literature or what that means doing the experiments yourself Doing you know many many repeats at many many time points to get the numbers So all of these have meant that models have not really in my experience gone very much beyond a certain scale When you're dealing with detailed mass action kinetics Okay, so this is where we thought we would try something new Sanket which means signal in Sanskrit This is the signaling and neurophysiology Knowledge resource for expand experiments and theory. This is a horrible background, but that's okay We we like them. So this is now We've just got a prototype website up which describes which embodies the workflow that I'll be telling you about It's we've had the consortium has been in existence for some Some months now for getting close to a year. We've had a we have a regular newsletter We have members from all over India and in fact some from other places and we'd very keen to get other people involved as well So since this is a session on workflows, I'm going to focus then on the workflow that we have in the Sanket project So we start by picking some process some cellular process or neurophysiological process that we're interested in So let's suppose it's something to do with synapses and their plasticity Once you've picked a process, then you want to make a model of it You want to say that these are the mechanisms that we believe in We want to get some data to populate those models and then this part is where I'll focus most how do you make the model as reproducible as accurate as good a representation of all that data and And then once you've got the model, there's many many things that I'm sure you all know you would like to do with it which range from Making predictions about how the brain works how plasticity works to examining various kinds of disease and Diagnostics, I mean there's a lot of things you can do with a nice detailed model of cellular function Okay, so to start with We were all in different labs interested in different things So we went about picking processes based on things that we were interested in one of them, which I'll be talking about most is the autism that is a project to simulate autism So that is what I'll be telling you about in most detail But there's other people for example, so we thought not currently who's from ice I said Pune one of our institutes in India She's very interested in presynaptic processing And there's another grouping of people who industry in synaptic plasticity. Basically, we can expand the topics We would like to look at depending on how is interested in it. So that was what we picked the members of this group looking at the autism project are Listed over here and they're from different parts of different all over the place Frat the autism as all of you know, I'll just Remind you very briefly One of the major causes of it is is is fragile X syndrome There's what happens is that there's a nucleotide repeat which causes a little fragile dangling bit on your X chromosomes and This can lead in a fair fraction of cases to mental retardation the protein. That's been implicated fmrp has effects on translation a fair number of Dendritic mRNAs are in fact targets of this and it leads to different kinds of morphology in the Compared to the wild type the mutants tend to have long thin immature looking dendritic spines so this is the system we decided to look at and in particular we decided that we would look not at the Systems or cellular level. We look at the subcellular level. What is happening at the synapse? Where how is it that incoming synaptic activity is converted into protein synthesis? So that was the system that we picked and of course the other groupings picked other systems are the sets of pathways to look at So how do we go about making a model? What we did was first of all, we need the model Modeling framework in this case. We chose the simulator moose, which is partly because we have developed it But also partly because these calculations Certainly involve multiple scales of functioning ranging from the signaling pathways the molecular Calculations that I'll be focusing on but also looking at cellular biophysics and also because there's a synaptic morphology phenotype also on how the geometry of the cell changes To do this we dipped into a database docs, which has got many many signaling pathways Particularly to do with the synapse We took a model of synaptic kinases We took another model related to activity driven protein synthesis and because we are looking at autism There were certain specific pathways that we know from the literature that had to be incorporated so we took all of these things and stirred well and mixed up to get a really really big nasty model Which is the so that is the thing that I started out showing you So it's got the additional bits in there and we merged together the kinases and the various pathways involved in in Protein synthesis driven by synaptic activity Okay, so this is a big ish model. It is substantially larger than anything. We have tried before I know that some people may have tried yet bigger models But let's see let's see how we go along with this because the goal of this The next step of this as you will see is how do you make sure that this model has actually got his feet on the ground? That it's got some basis in reality Okay, so just as I just to reiterate each of those things is a fairly complicated block of reactions And we need to parameterize we need to specify numbers for every single one of those reactions every single one of those Concentrations has to be defined in some way or the other so that is our challenge And of course, that's not the end of our ambitions at some point We would like to include for example reaction diffusion. We'd like to include the stochastic processes That are taking place in the communication with the spines and dendrites We are going to certainly put in electrophysiology and calcium dynamics and morphology change But for now we have enough on our hands as I'm sure you will see in a moment Okay, so that was how to make the model Next question was how do we get the data now in a perfect world? We would have entire institutions at our disposal generating these numbers for us This is not so easy, but we do have some Activities going on in fact we have been able to tap into a consortium activity on our own campus Where they're working with human IPSC is human induced pluripotent stem cells Which means that basically you can take a wild type and disease Patients sorry wild type and disease cells Differentiate them into neurons and then study their physiology and chemical properties something you cannot normally do with human brain tissue So there's a bunch of things we have begun to do on them We only have a little bit of data from this at this point. We would like to get human slice We actually have a line into a hospital which Processes such things so we hope it in due course to be able to do that that we haven't yet gotten going Mouse tissue culture. There's a lot of things that we can do with that and mouse slice now as you heard Right at the beginning of the of the of the day that you know mice are not monkeys and monkeys are not people either and Certainly the the what we have in a dish is not really representative of what's going on in your head But what we the best we can do is to try and get a few Sample points and try to interpolate or extrapolate to figure out what might be going on in the in the actual brain in the human brain So this is how we're approaching it the other side of it is to go to the literature and the literature is vast and voluminous and Extremely hard to parse extremely hard to extract numbers from so this is a job that Nisha in particular has been involved with She has so far curated Some 230 to 250 experiments Which I'll show you about in more detail and these are going to be the fundamental constraints that define our model so and one of the tools that we have in the in the Fine sim project in the in sorry in the Sanketh project is in fact a way of taking Experiments from the literature and sticking the numbers together Putting in as you will see the the aspects of experiments that need to be replicated by the model Okay, so that was the first two steps of the workflow now We come to the heart of it Which is how do you once you have this model once you have these experiments codified how do you get? How do you get to improve the model? So They're various components for that. There's of course the simulator you need to be able to run the calculations then there's fine sim, which is a project that I actually spoke about In a previous meeting in I NCF meeting And which has been published. So what this does is it allows it to it specifies a way of codifying? experiments and To take those codified experiments and run them on the model So you take the stimuli that were given in the experiment and you apply them to the model and then you compare the model outcome with what happened in the real experiment and you can do that from the web interface which can also manage this Growing database of experiments and then we run all of this through something we call horse horse hierarchical optimization of system simulations and What this does is it deploys the fine sim numerical engines the calculations of fine sim which again lie on on moose and then it can run these things in parallel and It deploys these to do the optimization using some standard algorithms, and I'll be telling you about these different steps Okay, so fine sim is this framework to integrate neural data and signaling models It's got its own website, which is all part of the SunKate portal and the approach is as follows So here's our model which I have described to you For the experiment we need to take the stimuli that were given what were the prop? What were the conditions under which the experiment was done and we also need to have what were the readouts? What were the outcomes of the experiments? So we put all of these things into a database we have a database of models We have a database of of the experiments and then we run the calculations on it And we compare the simulation output with what the actual experiments gave us and we can then compare the The real versus the the simulated results and use that as a score to decide how to improve the model So this is the flow chart for Doing the calculations within fine sim So there's a whole bunch of experiment types that fine sim now accommodates You can have the most straightforward one is you have your your system of Cells or or Reactions in a test tube and at some point you put in a reagent and you get some kind of time course of response of some readout Another very common thing which is used a lot in Biochemistry experiments as you do a dose response curve you apply fixed amounts of some stimulus and you measure how much response there is bar chart Experiments are common You can also do things in the electrophysiological domain. You can do current or voltage clamp experiments. In fact, this is Sort of under sampled version of the Hodgkin-Huxley Paper from 70 years ago You can do standard Potentiation kinds of experiments and many many more so these are all codified in in the in the in this workflow and Now we can codify it on the web You can run it from the web and in due course very soon. We'll be able to optimize it also from the web And I'll tell you a little bit about that Okay, so let's get to the optimization part which is the the heart of this So level one is we need to optimize sub parts of the model and This is something which can be parallelized because you can do each one At the same time you don't at this point We're just separating out each of these 40 or so different signaling pathways and you can do the optimization of them individually So that's nice Level two you want to take these modules and look at the crosstalk and so we need to optimize for the cross reactions That is a little bit harder to parallelize But some aspects of it can be done and level three we want to do multi-scale calculations. That is we may want to Optimize for when you have certain reactions occurring say in the dendritic spine or in the spine head and others happening in the dendrite So now you have to cross look at the crosstalk between those and so this or for that matter between the electrical activity and the chemical activity so each of these things can happen in In in a hierarchy and the reason one other reason why it's important to do it in a hierarchy is because if you have if you have Multi-thousand parameter model and you try to optimize all thousand or so of those at the same time You're not going to get very far Those of you who done who try to optimize things will realize that even the best algorithms will run aground if you try to do that so Doing it in a hierarchical way, which is also fortunately how one does experiments Makes it feasible to do this Just a glimpse of this So you might be able to do a reasonable optimization for a small part of the model, but eventually you will need to fit in the whole model If you start with a single subset say any one of these and each of these has now been optimized You would take the individual reactions and you would take a bunch of experiments which happen to Probe those reactions which constrain the rates of all these possible reactions then once you've decided the Reactions once you decide the experiments you figure out which parameters you permit your optimizer to tweak and those the ones in in red and Having done that you you unleash your optimizer, which is you know a fairly standard algorithm To optimize these you can assign different weights to say which experiments you want to assign more credibility to and then you turn things Turn the handle basically you let the the computer turn away at this And run the calculation. So basically what you're doing at this point just envision it you're running You're you're specifying an experiment. You're running the experiment on a model You're comparing the experiment output to the model output and you're giving it a score and you're doing this in parallel for a whole bunch of different experiments which all engage these pathways and each of so having run through all of these experiments you get a sort of summed score Saying that at this point the model fits this model fits all of these experiments to this degree Once you've done that you can go to the next iteration of the optimizer So all of this is done can be so all of these things can be done in parallel And then you go to the next step of the optimization process And so at the end of it you end up with some modified parameters You end up hopefully with an improved model and here's some examples. So this is where you started and This is so this pre optimization post optimization. So this one of the experiments Not such a great fit a much better fit. This one started out really terrible fit. It's still not a great fit But it's better than it was This one actually got worse But this one got better. So the model is Approaching not just one experiment, but it's trying to approach at the same time all of the experiments, which is why it's a somewhat touchy optimization problem and So in this manner, we have actually gone through the process So I should say Nisha has gone through the process of optimizing all of these individual modules Independently and the next step is to look at the cross talk. So this is where we are right now with this process of optimization and It's a similar kind of thing you pick up a set of experiments that define how the pathways interact and now you try and Try and optimize for those rate constants so We've recently had a very productive Google Summer of Code Interaction student being Hao Chen from Peking University He was a sort of Handheld by Surabit who is from from our group and so this I think is a kind of a nice illustration of the of the reach of the INCF that is we spoke to the neuroscience gateway people I mean Majumdar and his colleagues We had Google providing input for the Google Summer of Code project and of course the INCF brought it brought it all together Okay, so this has been the the the core of it how we went about the optimization I'll give you a glimpse of some of the things that we are already beginning to do with these models One for example is to compare what happens in the wild type version of the model with various mutant versions of the model So we can optimize based on Mutant parameter mutant experiments on mutants to get a different version of the model that applies to the mutant So for example, we know that in the mutant the levels of these different molecules Have been changed and we can run the calculations and say that for example for certain stimulus We get different kinds of responses. So this is a very early stage of Analysis, but we can already start to play with the models and say this is what happens in the wild type And this is what happens in the mutant Another thing that one can do is one can play with or examine how well the large body of experimental literature fits different theories of These mutants so for example, there's the Metabotropic glutamate receptor theory in blue our theory of a fragile X syndrome, but there's a competing theory also the cyclic ampi theory and Which of these is probably accounts for more of the observations? This is something we can now begin to play with and there's of course for pharmacology, there's the Enduring question of what are the best targets and what will be the side effects? So for example, if you were to believe in the m blue our theory, you would expect that this would be a useful target But then you would have all of these as potential side effects so we can calculate those Or if you believe the cyclic ampi theory, this might be one of the targets And then you'd have side effects in different parts of the model So these are now things we can put numbers to and say these would be the side effects if you put in a drug Which targeted that molecule of that molecule? Okay, so to wrap up then What I've done is I've just tried to describe to you the workflow that we've developed for this project This consortium it starts with coming up with a system that you're interested in Then one has to develop a model for that system of course closely based on a lot of data that pertains to that and then we have a whole System a whole workflow to take the available data Structure it and be able to use the data in a very systematic in a principled way to make a better model and Finally then it's up to the to the people in the project to figure out what they want to do with it so that is the the essence of this project and To wrap up I would especially like to thank Harsha, Nisha and Surbith who've been core to the project Jyoti, Vinod, the various other people have contributed their various collaborators and consortium members and funding from various people. So thank you Fantastic project. Congratulations indeed So far I've seen only big models which have been done within this ecocyc and the biocyc projects Which are the microbes and really small organisms They are still open systems, but it seems to be a bit easier to handle this openness than in your case So what's what do you do with the rest of the world? So to say when you have a model like that? So I'm not so familiar with those models, but are those more the gene regulatory networks? Yes Yes, yes Right, so one thing is that they have access to I would say a much more structured set of experimental data because you can sort of grind through those experiments in a You know a very systematic way you have the mutants you have the metabolic conditions, so that that really helps This is I mean certainly they've done an outstanding job on doing this But I think here we are up against a much more heterogeneous set of experiments And that was one of the big challenges that we had to face to come up with a way that you could take Very diverse kinds of experiments and still use them But the question is about the openness right because in case of microbes and small systems You know what kind of nutrition sir molecules are interacting with this system in your case. You're just a part of it It's true. Yes. Yes, so so how do you treat the openness? That's yeah, that's a great question. So this is this is the question you ask the person who's looking under the lamp post for the key Right. Yeah, you're looking there because that's where the light is. No, it's based on Our best judgment that is we take what we can see from the literature that these are key pathways that must be involved Which is why we incorporated a few additional pathways beyond the ones that we had started with It is very likely that as we go through this process We will say that okay, some interaction is simply not working with the available model So we have to extend the model. So that's a systematic way of doing it We'll know that is that is part of the part of the process any further questions So actually I would have one question just So I was wondering like these that the advantage of having these workflows is is that you can actually rerun? Yes, these workflows again and again. So from your experience What happens if you get new data and you get to optimize against an extended data set? How stable are these estimates in your experience so? at this point So generally, okay zero order is reasonably stable Every so often new data will show us that our existing model is inadequate In which case we need to go back and add some more interactions that are not there Currently, but the whole point of doing it this way is that you can incrementally add new data as it comes in Okay, if there are no further questions, let's There's one more in the find some workflow. I imagine there's there's parts of this that are very Specific to the kind of system that you're you're you're optimizing and then other parts that are more general right that could apply to maybe some a different kind of Experimental system or a different level of modeling to what extent are those Are those are those make like is there a distinction between him between those or is it kind of built end-to-end for this this kind of these kinds of models? so it's It's it's extendable The things that we put in were based on our general interest in looking at neurophysiology and the underlying signaling So that's why we have a range of things from the standard biochemical readouts to also the standard electrophysiological readouts and the slice LTP kind of readouts should more things come up that need to be incorporated we can we can do so the Commonality in all of these is that you specify the experiment you specify the experimental result You run the simulation and you can compare and therefore you can apply the same algorithm so the specification of the experiment has two parts you specify what was the Preparation so in other words what part of the overall model was necessary was used by the experimentalist and You specify what were the inputs so let's say a stimulus at such a time or a shock pulse at such and such a time and Then you also specify what were the readouts which could be a chemical readout or an electrical readout and That's the specification