 Alright, so this is being called around table that it, it's the same square table that we've been using for all the discussions. And we're going to lead some discussion about what's missing from our understanding of genetic architecture across levels of biological organization. And so our, our panelists have been thinking about that in preparation for this meeting. And we'll just arbitrarily start at that end of the table and work our way forward, and we don't have anybody online, everybody's here right now. A generous of you Nancy, thank you. So, I'm surprised that during this meeting we haven't talked about the different between prediction and individual effect. I seem to be like, one of his themes that like front and center right many of the conversation we've had about variance explained, and depending on what we're looking for, like this, some of the questions that are posed about. So the hierarchy of the causal flow from cellular effect to tissue to organism. I really call it by whether or not we are trying to, what are we trying to understand what's the question we're asking I think in this room right we care about say different thing right. I think it's very explicit about what he's looking for. And I think many of us, I'm sitting complex architecture for very different reason I think I don't put words in your mouth Peter but like, your work as centered from estimating variance and processes but many of us in this room care by individual effects right, and, and the perspective there is very different. And they're also touched on this G by a question that we spend most of the time is during this morning right how individual that curiosity alleles, and how busy lives themselves evolve as a consequence of history call exposure potentially shape how. And I think that's this question of prediction versus individual effect, I think it's very important I hope we can talk about this as a panel and as a group going forward. Yeah, I think that's a great point the one thing I want to go back to I think something that I'll get that earlier about physical models, I think we talk, we talk a lot about today about this idea of generating new data and they have enough data to answer these questions. And I kind of want to, you know, poke a little bit of saying that maybe we'll have the right to school methods either to think about some of these questions particularly across multiple scales. So I think about ways to integrate information across those scales. And so you know you think about, you know, do I have models that will explain something, things I think about on the, on a SNP gene pathway network like level, but the you know, do we have a theoretical frameworks that also are in place for us to as a foundation to build those methods on. And so we can talk a little bit about some of those questions as well. Hi. So, so we've, we've got prediction and methods and now I want to talk about phenotypes. So I think in the previous session there was several people who brought up, you know how how come we haven't talked about like definitions of, you know, either diseases or phenotypes or what is it that we're looking at. And in at different levels of biological organization. It is expected that those that are closer to genetic facts are going to have like, you know, much more explainable genetic findings and, and one of the, one of the questions that you know always is on on the thing now, which is you know, endophenotypes or molecular phenotypes have surprisingly complex architectures. So, so there's a several things to think about that one is, are we looking at the right endophenotypes, if they are so complex maybe they're not what we're looking for. And also, you know, if, and this brings up all of the other things about, you know, plyotropy and and polygenicity and multiple things having conversion effects on a single phenotype. So we could talk more about phenotypes in this discussion as well. Yeah, so I don't know how much sort of new new stuff I have to add but I have really loved a lot of the conversations that have been going on. So from this question of what's missing from our understanding of genetic architecture across levels of biological organization. You know, there's still data missing and especially sort of comparable data sets so we've, you know, talked a lot about and one of the questions is about this idea that maybe intermediate molecular phenotypes might have different genetic architectures maybe simpler genetic architectures. I think that often the inferences that we're making about that are coming from studies with very different amounts of power across very different sort of types of people and I, you know, I think that we're all converging maybe on sort of basic principles but I think that it would be sort of nice to get us all talking about whether we really do think that we've, we've nailed down the fact that there are different genetic architectures across different levels of, of biological organization. I think also across different scales of biological organization we're often thinking about different sort of mechanistic and generative processes so at the, the level of molecular QTL as we heard about this morning. We often do sort of have more of a mechanistic biological model about this thing binds this thing and it changes gene regulation in this way and I would like to see you know more connection of that we all like to see more connection of those sort of mechanistic and generative processes with some of the larger scale levels of biological organization and I think we're all interested in how we get there and I'd love to sort of keep that discussion going. So first of all I'd like to thank the organizers this is awesome being here and being part of this fabulous day. You know I'm really excited about the next 10 years. Because I feel like we're entering a sort of post discovery I know we're still going to be discovering but now that we've got hundreds of variants for every trait I think it's time to start really doing some really interesting biology with them. And to me the major question is, what's the variants of a lillic effects themselves so so we've talked a lot this morning about variants across contexts. But even within a context, there's a really big question as to how variable is an effect for an individual with with the same genotype. So two weeks and I think probably functional biology can inform that. So what I mean by that is things like eqtl give us a different perspective on what genotypes are doing. So one extreme you have a model which I call the sort of kaplunk models kaplunk this game where you have these sticks holding up marbles and the aim of the game is to pull out the sticks. And that's the point where one stick comes out know the models fall through. So the analogy would be disease would be sort of held up held at bay, by all the sticks which are the G Y sits, and then for each individual iteration of the game, one particular stick is the one that really matters so in that model individual effects are really very very small for most people with with the snap, but really big for a particular person who had me had a particular exposure or just by chance that was a big effect. The main is that like effects are really the same across for everybody everybody with an extra allele of this has an extra millimeter of height, and that's uniform. So addressing that question to me is critical and the way I think about it looking at eqtl across context. So when we did our work in Morocco a bunch of years ago, we found hundreds of eqtl in the blood. What struck me was that there was absolutely no G by e for individual a little like effects, even though the different environments had really different means of expression so there's just parallel differences so each genotype was the same effect, even though the magnitude of the absolute levels of expression were very much culturally context dependent. And if you think about that it actually means most individuals in any group with the risk alleles have pretty much the same expression of people with the protective genotypes. Okay, so the vast majority of people have actually the same gene expression. And there's only extreme individuals in one of the environments and so that sort of sets up the possibility of genotype environment interactions at the phenotype, you'd actually don't observe at the molecular level, and raises the idea of what I call common variants of rare effect because only a few individuals with that risk genotype are the ones that are extreme. So that's the conception that's sort of one way that we can think about using molecular cell biology to truly study the relationship between genotypes and phenotypes by asking, you know, what is the effect of the genotype on the biology of the gene expression trait, and then linking that to cellular phenotypes. So I'll stop there with some other thoughts. It would be great to, so everybody should feel free to respond to the issues that others have raised, but also to address some of the specific questions. You know, we've talked a bit about plyotropy and how the same phenotype like age might sometimes be a cellular context might sometimes be a much more organizational context. But what is it about, you know, is it really true that that intermediate. How is it what is it about phenotypes that drive the complexity and potentially the, the challenges of understanding genetic architecture. Phenotypes like BMI or major depression, maybe have a lot more opportunities for input from environmental exposures, then we even know and understand. And so those are, that's another way to think of complexity with respect to the phenotype. There's some important questions raised about how we think of the analytic models and, and how the scales may be different across the genetics and non genetic exposures, and, and more, more ways that we need to think about better modeling. And, and the dynamic processes so let's let's go back through. People should feel free to respond. And then, so after that we'll, we'll open up for more questions and discussion. Yeah, so I just had an extra thought there I guess temporal dependency of a little effects is I think really really critical and particularly in terms of the buffering that word was used this morning. And I think that's a critical element of effects over the life course or body exposed to disease response to that disease by changing gene expression. And I think that's a critical element that we cannot get to within vitro systems because it's just the whole body the whole organisms responding. The extra perspective I have on that is whether or not, you know, under persistent stabilizing selection you can actually evolve systems that are robust so so technically I use the term canalization and it can't mean to evolve robustness and evolve stability. So is it possible that that that actually persistent stabilizing selection leads to changing the genetic architecture in such a way that you tend to promote that stability and buffering. You know modeling done by people like Sean Rice going through and Andreas Wagner and others is putting it in that regard. And so maybe going back to some of the questions that are up here and especially this ongoing discussion about pyotropy. During Naomi's talk I was wondering sort of whether we could use some of the interesting observations about pyotropy that we're getting from to us like the fact that there are some of these variants that seem to break the relationship between minor allele frequency and effect size scaling and sort of go after those functionally and try to learn about them, given that they seem like they are probably really important phenotypically that they are going to be affecting some things positively some things negatively, and sort of sort of work backwards from things that seem to break some of the sort of common GWAS patterns or some of the common ways that we think variants in complex traits should behave so what are the outliers maybe we can do some functional experiments with them and try to learn more about their biology. So I want to respond to something what Nancy was talking about G by just now some environments. Well, depression being one of those things that is affected quite a lot by, for example, early life stress and stuff like that. So, so I think in the morning, there was a few discussions on on how linking for example G by results with T was results that would be able to link for example environmental effects to complex disorders. And one of the things that is that I think in some discussions that we have like in coffee breaks and stuff like that is, is whether those things are on the same scale or in the same direction. So, for example, if a environment affects a gene expression, is that going to be exact, well, a higher gene expression is a higher gene expression but the environment affecting the gene expression is that the same thing as you know, genetic variance like affecting the gene expression and and would that have the same outcome on the on the on the phenotypic level. So those things we could think about a little bit more when we were thinking about you know, modeling G by all the way towards complex phenotypes. So I'm going to go go back to what I was where I started with the expectation I think that many people I mean that I had was that when we look at cellular phenotype but you can tell the cell level. And as we go up the hierarchy, based on this question like things integrate and maybe there's more buffering as we go up right because it's the other ways that we sort of penalize in some way. There's this complicated trade off where there's so much polygenicity also in this big pickle right where like, like your question about how do we do biological experiments like what does it mean is era of such vast polygenicity. Like, I, like we've all been there, we write a paper is polygenic architecture so if you're asked us to validate the effect of something. There were two winners that was like someone swing on the top. We don't know, we don't know much more right if it's polygenic by definition there's a ton of effect that adds up to something that account for most of the violence. And so how do we begin to integrate this both the complexity of different layers, and the this vast polygenicity into biological experiment actually teach us something beyond the prediction for prediction is simple right I mean it's not simple sorry it's very complicated. But we can all understand that we have some framework to understand how to make a prediction in the black box of biology. But if Andy wants to make a drug based on this sort of polygenic group. That's going to be very hard right and I, I struggle with integrating these layers, and then you had aging and you had, like, all this, this thing that are going to modify this buffering system that are robust when you're young and not robust when you're older, these polymorphism are going to be cryptic sometime bringing it out. So I think these are some of the challenges that I think are hunting and most pressing. I want to kind of add to that because I don't have a wet lab so thinking about how to validate in silico across scales also something I've been thinking a lot about. I don't have a lot to add outside of like, you know, you know we build models all the time, you know, we might think we have a story on one layer that we find another story on this going up in scale on a different layer and that complete contradicts a lot of fun on the previous leg. And so you know one thing I've been thinking about you know this is, you know, how do we come to more cohesive, you know, understanding our results in our in silico experiments when we think about, you know, is g by playing a role here or is that some kind of confounding or bias from from something else that's something like, for instance myself and our goals back there somewhere I've been thinking a lot about recently with certain papers and and so yeah that's a huge challenge as well we think about these different like multi scale type of approaches. I feel like a lot of not, I think it touches a little bit on what you were talking about just now like on the integrating different layers and polygenacity. So, so I think a lot of work has been done over the past like decade or so to sort of identify some annotations or sort of tissue specific gene expressions and people have been using that as a unit for defining like things to test basically, like against diseases. So for example, you know is in the region g was hits more enriched in this type of like genes expressed in this type of cells, or something like that. And maybe one of those, and that gets us as far as you know, the cell types that have been annotated. And whether that is the right level of organization that we should be class like clustering things to get enrichments on is it's hard to say and when do we know whether we got the right pathways. So that's something that I think is one of the biggest challenges right now if you're hoping to to identify those polygenic pathways and and to use it for example for stratifying heterogeneous diseases, etc. Yeah, I think that that's a great point. And I think also, you know, as we've talked a lot about this morning that the those sort of annotations that we're usually using about where are these genes expressed what cell type are they often expressed in that those are all, you know, baseline healthy state healthy people, and that, you know, I think a place the community could go is some sort of more large scale community resources for thinking about, you know, developmental trajectories cells in different states, disease states, different environmental states right I think that would be helpful for annotating all of these, you know, polygenic genics of complex traits lists that we have and thinking about their biology. I think we kind of need a poetic naturalism for genetics so this is sort of Sean Rice sort of concept for Sean Carol's concept in physics but but are there emergent properties at the organismal level that we that we wouldn't see at the settler level. This is actually to wonder if now can expand on her comment that linking G by e to T was so how do you see that happen because obviously T was does not capture environment. So what's the best way to go about that would be are we better off investing in well okay back up so we probably spent over a billion dollars on G was to this point. Maybe $10 million on the sorts of functional assays Francesca and truly we're talking about. So if the NIH will invest in this sort of research to sort of bridge from functional annotation. So all the stuff that encodes done, really bridging that now to population and quantitative genetics. And you would say well, where would you spend $100 million would it be on cellular assays would it be on more and more screens of more people across environments, which will be restricted to blood I presume. Are we doing more CRISPR assays. I think that's a really important question to address and answer the, you know, are we limited by data, are we limited by theory, those those sorts of questions. I want to draw people's attention to a different kind of perturbation that maybe has some things to help us learn from. The shocking thing to those of us who saw the GLP one our analogues, you know from very early days, you know, come out and be used, and see how ignorant we actually were of, of how that biology was going to work. Right. So, these are drugs that were designed, you know, followed up. I would say the targeted impotence should improve insulin secretion should maybe lead to a little weight loss that was those were the early ideas. And the magnitude of effects for those drugs is shockingly greater than what what was envisioned by the people who developed those drugs. So these drugs lead to much more weight loss than was expected, and through mechanisms that were not really imagined so much bigger hits and reward system biology because they're also expressed in the brain not just the gut, but also improved cardiovascular risk profiles, and hugely improved. I mean in short trials, fewer cardiac events. I mean, a big deal for cardiovascular health that was not anticipated at all in the development of the drug. And this is telling us really important things about how these systems are connected. And, and, you know, it's, it's back to the pliotropy but it is. It's a drug perturbation potentially that helps us grab hold of bigger pieces of biology that are connected that we didn't know how strongly they were connected, and that we should be able to use in thinking about designing experiments that help us understand these sort of cross organ. Unexpected shared architectures unexpected communications. And that the. So the GLP one R is is one thing but the sglt two drugs, whose primary activity is just peeing out more glucose are having similar surprising effects and improving cardiovascular and kidney function. There's so much we don't know and understand about how by these layers of systems talk to each other, and the question of whether there's experimental things we can do like now really looking harder at at drug perturbations of systems, these drugs that are way more than we expected. In hindsight, could we, could we look at relook at them, the animal model data, and actually predict this, or not in in retrospect, would we have put these systems together and the way they, they seem to be put in thinking wrong about diseases like diabetes as anything other than just part of the whole cardio metabolic system. Super system. So I, I think, also, let's now start opening up to two questions from the audience. Go ahead. So Michelle Nevar to answer them. My question about the GOP one thing and it's a question has been on my mind about many things we discussed today is like, as an epidemiologist my intuition goes to that that's probably like 60% the weight loss, improving the heart conditions, just like the action is mediated through the weight loss. To some degree, it probably goes a bit too quick to be entirely mediated by that but like we all have different intuitions about the causal chain at different levels, and they interfere with each other. Because while there might be a biological effects of GOP one on the heart condition that you may be able to figure out in the model system. There is at the same time massive weight loss that just has like a biological effect. And that's just that's still a reasonably simple thing throughout the day we've been discussing sort of all these systems, or how we could study parts of this system but at the same time, there is causal action at every level of biology all the way up to sociology. So how are we going to try and how would you how do you envision us trying to entangle those levels. Well, there is emerging data from clinical trials that the improved cardiac biomarkers way precedes the weight loss and is not accounted for. Meaningfully by the weight loss I mean it's way beyond what you would expect for the for the magnitude of weight loss in terms of especially the timing. But great point that BMI is is a super driver of medical phenome all over the place and and anything that's going to affect that will have have lots of downstream effects and how how we disentangle that is very hard. So I was just going to add there's another example that's probably not right. Well, maybe it is related to my but in the opposite direction is long COVID effect on heart diseases is also systemic somehow in ways that you cannot predict from this creation. Are you have a question. Yeah, I'm. Thank you. So, I think in my mind the question is whether you know many of the common questions we've asked throughout the day. We've just asked of every phenotype. So far, you know we've studied phenotypes depending on a personal curiosities, or whether a set of, you know, not everything is a model system some model systems have been better than others for what the questions were. Whether a certain set of phenotypes may teach us more simply for whatever the scientific, you know imperatives are. And I think we're asking very, very hard question that I find I think the only conclusion I came to is answering them in the absence of a particular phenotype or disease is going to be very difficult, and it very well as possible. Some phenotypes say having to do with escrow disease is far more answerable than depression genetics may not be at a stage we couldn't study complex diseases 2030 years ago. And didn't seem to bother anybody. So why do we expect that we should be able to understand depression and what have we done to change the probability of understanding it so it may be helpful to hear whether there are specific questions and specific paradigms that would help us understand each of the questions. Just a random thought and response. I took a class in urban planning and urban policy last year just for the heck of it. And the professor started out by by talking about cities and the architecture of the cities and the street plans and how open streets were and what impact that had on crime and so forth. And he said, you know, there's no point breaking it down into its component parts because you never got to understand the city by putting them back together. There's nothing more complex in the city. And I'm sitting there thinking, well, brains pretty complex immune systems pretty complex, and we're breaking it down to its component parts, and sort of assuming we can just put our understanding back together. There's parts and their effect sizes. And so his starting point was, there's no point even trying that we can just take little bits of the city and understand that, you know, better sight lines prevent crime. And, and we don't do that. So that's again coming back to the question of emergent properties and thinking about that. So, so this is Aaron Pinovsky have a question back here in the back. My question is basically, how do we even think about doing these kinds of studies in a changing world where there's changing rates of changing into sentences of disease changing sort of environmental context in the broader sense when we spent the day sort of we don't really know what an environment is. I'm just thinking about something that might be even seen as relatively tractable like addiction or something like that but we have, you know, a vastly changing landscape of which people become addicted which drugs are available which forms of addiction end up killing people versus which end up being sort of chronic and manageable. So, you know, so we might think okay there's a genetic architecture genetic underpinning to addiction to opiate addiction. That's, that's great. But how do we, how do we even think about identifying what genetic variants are in a in a changing context like this right so if we did a GWAS in. Here's a maybe a concrete way to put I'm not sure if this is the best example but if you do a GWAS in 2023 and you did a GWAS in 1970 would, would it be the same thing. And, and if, if yes if no I mean how do, how do we then even think about what genetics is doing, or what the genetics of complex traits is doing with those sort of thought examples. That is a great example if you did a. If you could do a GWAS in 1970 for cardiovascular for myocardial infarction cardiovascular disease, you would definitely see much more of the LDL based component, then you will today because statins treat that component of risk for myocardial infarction so effectively. And so that's part of the distinction between prediction and and what we can learn from GWAS and, and how environments and drugs are going to, you know, to the extent genetics works, and we pick off variants that we can use to improve health. Things will continue to change, I mean, they will, they will be radically different. When you look at at GWAS of the modern data, you can hardly see the LDL signal, compared to things like LP little a because that's that's what's driving in people who are on statins. So I, I think, yes we have to be aware of the dynamic part of the world that we control but then of course there's the whole dynamic part of the world that we, we don't control and we don't understand and that's yeah. Just a small, small comment. So I come from the model system world I can flies and, and people can do experiment I can take the exact same population and change the environment and do GWAS on the same trade and look at how different the architecture is many people are doing this kind of things. And is, it is quite dramatically different right the amount of cryptic genetic variation that get revealed is substantial. So, to your point, yes, it's a moving target where I think that's what you're pointing at, right. And that goes back, as you just pointed out, what, what are we trying to understand for prediction doesn't matter right is in the context of that population at that time. But for the context of understanding the biological basis of the variation depressive behavior. That's a that makes a huge difference. And most of what we're seeing is changing over time and we have these dynamics that are dependent on like what heart attacks even are at different times and I was wondering about this earlier to in the last session but like how much of polygenicity is literally just time, like traits that take longer to generate or traits that take the develop over a longer period of time are just ones that are more polygenic. And it just so happens that there's a lot of these intermediate phenotypes that are short time scales, but most of them are most of the larger scale traits that we're looking at their organismal level develop over the course of many years and so as a result of that they have more time there's more environmental exposures, there's more variance as a result of the fact that the society and my culture that we're living within is changing, like how much of this is just capturing that. I don't have the answer but I think that's a great question and a great way to think about it. I think we also can't disentangle, you know, age and time from sort of accumulated environmental exposures right those are always going to be entirely correlated and a lot of the complex traits that we're interested in and these complex diseases are diseases of aging diseases that creep up later over time and we often can't really disentangle sort of biology being different the body being different things breaking down over time with just you've now lived for 70 years, experiencing many different environments and your, you know your body is sort of up taking that so I don't know but I love that idea of sort of thinking about it as time. I'm going to echo Greg's question people can consider this a homework question because I think this is part of what Jen's going to come back tomorrow but like from, you know from the NHGRI perspective we're often thinking like, what is, what is missing that we could help fill some of these gaps. And two things that I'm hearing that I'm interested in one is like just like, are there resources missing or in certain, you know, are we only, you know, the, it's at lampposting are we looking at the places that are well annotated and is that biasing what we're finding and you know is some of this coming up with less biased annotations, but then to come to your point Greg. How do we get those integrated views right if we want to be not just looking at the parts but coming across systems you know it's something that's equivalent to sight lines. And what are the things that we could help allow those types of views to be better, better seen or better pulled out in versus what we currently have in terms of existing methods existing data existing resources, because I don't have the answer to either of those questions so I'd love some input. So looking at it across populations, my command is doing is great but then that's complicated by ancestor. Yeah, I mean, I'll definitely sort of plug that I think that even when we're thinking about the environment that we are often thinking about the environments that we experience in the US or in sort of in the global north and there's a much larger range of environments that are true across the world. In some cases that is going to be, you know, confounded by ancestry but I think that, you know, larger scale, sampling larger scale incorporation of diverse environments and including environments that are more similar often to those in which humans can relate from a sort of thinking about the evolutionary history of humans the environments that we are currently asking people in are very weird. They're very novel relative to sort of what the body is seen through time. So I do think that there's more that we could be doing in that space. Yeah, I had a question. Oh, sorry if you're going to go ahead now. Yeah, of course. Yeah, no, I have a jumble of answers to various things that I was raised earlier. So I feel like our vendors comment just now about like different diseases being at different stages of investigation and we understand very, you know, different amounts about those diseases that seem kind of or the same mag well amount of genetic data we that we seem to have accumulated over the past number of years. For example, you know, mental health disorders versus for example cardiovascular disease or other more biologically understood things. Not only is it that we don't. The more complex phenotype is is less well defined, and the genetics is more complex we also understand the neural neuroscience, not not as well as, as all of these other things like the brain is. We do understand less about the workings of the brain that maybe off, for example, certain other systems. So maybe that also needs to catch up a little bit and and one of the maybe one of the goals that geneticists could, you know, aspire to is, is, can we potentially hypothesize some of those models, like let's say, you know, we're not neuroscientists we can't, we don't actually know what's going on we can't experimentally test a lot of things. But let's say we, we can come up with a hypothesis of how things work and then we can test it using genetics, we probably will be wrong. But then at least we know that that is wrong, and then we can go on to test the next thing. And, and if we kept doing that, you know, in a reasonable way then hopefully hopefully we'll get closer to the to the nature of the disease. And that's, I think, beginning to happen with the mental health disorders. Using genetics as a tool to test whether our disease phenotype is actually right or not. And then the other jumbled answer is, is, I think Greg asked me just now about G by E and Nancy said something about put the perturbation of a drug that seemed to have way larger effects on more than one phenotype. And that could be just, you know, that also ties in with something that I was thinking about earlier whether, you know, perturbed gene expression levels or perturbed molecular phenotypes are those the same things as I think that was what I tried to explain but not very well that's why cause confusion. Whether that's the same thing as genetically regulated molecular differences between people. And, and that's both in terms of magnitude and in terms of like how widely affects. So maybe, you know, if we could, if we're doing large scale G by your large scale per G by treatment or G by other perturbations. If you identify those molecular phenotypes that is affected by this, and then that could, for example, be the unit that we want to test things going forward. Yeah, that's how I want to add. Yeah, so first I'm going to this idea of like, you know, diversity on different axes and thinking about those different definitions of diversity, not even just ancestry but you know, other ways we can think about stratifying different populations. When to answer this question about like how to think about investments, you know, I come from a machine learning world. So something that we're doing right now I think everyone sort of term foundation models, which I think is like a kind of interesting idea which is, you know, this idea where you have a model you train it on something and is able to generalize and make different spaces. Foundation model really good in language, because language we see in our context, and I think what I learned today is that we are nowhere near that in this space right now is infinite many contexts but in language, it's very simple if I take a punctuation mark and I, and I move it like a comma to a different part of the sentence I know exactly how the syntax of the sentence change, right, and GPT and these other these other methodologies are very good at generalizing that because it understands all context, whether you ask a certain things and certain tones and certain ways, it can infer that even though I haven't seen it, right. You know biology we're just not there yet right we're thinking today we just don't have enough of these of contextualization is to kind of know if I change the comma, right, quote unquote in the sentence like how things may generalize. And so I came today thinking like oh maybe I'll figure this out here, but now I'm kind of like, maybe we should start somewhere to answer your funding question like this idea that you want to fund this I, you know, you know, perturbations enough that allow us to generalize to other places. And so maybe that's another way to kind of think about it and you know Microsoft has nothing to give away Microsoft secrets or anything but Microsoft is thinking about this idea of like what are scaling laws for different modalities in different application spaces right. So you think about like how many tokens of words or tokenizations of words or I need in a sentence in order to be able to generalize across many different languages, you can think about this, you do this also with sequences and proteins you can think about this also in terms of whether you're thinking about certain g by g g by e across different populations and maybe that's what you should start. So, I was going to ask on the topic of like polygenicity which we've been circling around the omni genetic model has that idea and it of the core pathway and the peripheral pathway and hasn't come up yet today and I was sort of wondering about that. It has maybe some connections to toolies idea of like points of convergence maybe is that sort of core pathway. That'd be interesting to explore that a little more and also in the context maybe hierarchy of, you know, core pathways maybe at different hierarchies. But I think I love that question because I think it also comes back to one of the things we talked about this morning with respect to whether the ways that the biology of environment non genetic exposures work does it layer right on the biology we know, generally, or, or is, are there way more ways that that we might learn biology by by exploring those things. And so I, that's actually one of the things I'd say in answer to Caroline's question, I do think we would benefit this as a society from understanding more about whether the biology of non genetic factors really will just reinforce biology we know, or at least, and maybe for more complex phenotypes with more ways that the environment can affect them. There are new kinds of bio, you know, new pieces of biology that that would be particularly useful and I, I, so I don't mean to hijack your question because I think it's, it's a great question all by itself, but it is related to that concept. And so that we discussed this morning. And, john, I think, maybe the. So, the core. Carmen what the word is core and peripheral regulatory networks in the omnigenics model. So that's normal. That's at the moment. I think most eqtl studies are done on non diseased cells, like tissues for like or non specified, not sure, not specifically by any disease there are disease data sets coming up now. So maybe one of the interesting things going forward to, to look I think is no one has a paper on this like some years ago about regulatory, I think he called it regulatory due to coherence or something like that. So maybe you know, in some disease states there are perturbed core versus peripheral regulation networks and that may be important to the pathology of the disease that might be an interesting thing to look into. One comment I wanted to make that relates perhaps to john's comment as well as to the prompt about genetic architecture across levels of biological organization is that, you know, with the tool of G was we're very good at identifying snip to end trait, or disease association so we can, we can measure very, very small effects, very reliably and repeatedly, but it's, it's extremely difficult for us to learn about how those effects are flowing through different levels of biological organization so so for example, if you think about, you know, somebody with a high polygenic risk score for some trait. What does that do to cellular states for example so so how do those, how do those effects flow through gene regulatory networks how do they change the, the states of T cells or you know what have they affected cell type is, and you know how do those affect tissues and an organism level function and like we, you know we can we can measure whole body types, extremely efficiently in huge numbers of individuals but it's extremely difficult for us to measure a lot of meaningful phenotypes at cell and tissue level. And, and I, you know, I think, I think that some of these things we can, we can try to get out with with experimental approaches like, you know, CRISPR perturbations in in cellular systems rather than having to do the same sorts of measurements in enormous sample sizes but anyway I think that this is a huge gap to, in terms of how to how to fill in the sort of the causal pathway steps from from genotype to to phenotype and you know sort of thinking about the ensemble of effects rather than you know sort of a one dimensional arrow from from from genotype to end phenotype. Thanks. Just to ask back to you Jonathan, if we, if we had a whole repertoire of IPSCs differentiated into lots of different cell types from people in far tales of polygenic liability distributions. How much of the biology of polygenicity. Would we explain at with just having the cellular information on on people in far tales and you know maybe mid ranges of polygenic liability to any disease and how much would, would we have to have this across levels of tissue and system to see anything I mean that's how that's part of the question that that is so hard to get out here that we don't have yet, but we might have pieces of it. Yeah, I think these are fundamental challenges we don't know know the answers to really so. One example, you know, our labs been very interested in and trying to do. You know, CRISPR perturbations of individual genes try to build gene regulatory networks, and then understand, you know, in a very specific cell type. What are the kinds of effects that variants that show up in GWAS have, you know, which parts of the gene regulatory network and most often perturbed can we identify downstream genes that may be having causal effects in that in the rest of the biological system. You know, I, I personally think that's a useful first step but at the same time, it is really only a very first step it's it's like getting one, one step away from, you know, so sissy Qt also linked variant to to the nearest gene. This is getting one step further and trying to understand how those effects are flowing through regulatory networks within the same cells. But anyway, I think there's an enormous amount of space for progress on this but they're difficult problems. Okay, um, I had my hair echo comments that the concept of time with what we're looking at, as well as sort of looking at a cityscape versus parts of it. And it was a really good example of how time changes things with COVID there was a GWAS of COVID. And if you look over the first couple years of the results they change with time, and they change rather drastically in some low side and that's probably because of more numbers, but probably because who was infected changed who was exposed changed, and then who sort of got hospitalized changed from various things and so you see a very accelerated version of this. And I guess, for us to think about in terms of levels of biological organization is that we have a pretty good idea of how micro we want to get in terms of the scale, but really how macro do we think is relevant to go to understand these different levels and how it informs the biology and the genetic architecture of traits. So I think it's a very sort of informative framework to think about and sort of necessary for us to move beyond even an individual and towards populations and societies and sort of how that influences even the basic biology of what matters. We have we have buses that we have to catch so we have to bring this session to a close. And I have the unenviable task of trying to summarize all these discussions today. I think this last roundtable came back to a lot of the things that were discussed earlier. So I, I don't think we need the big summary. I think we. There were a lot of interesting discussions related to play a trophy to selection to these levels of organization and, and how we need to improve our understanding. Also on the on the analytic ends, the the ways that our questions shape the models that we consider the way that the models that we've already built shape the questions we ask. So we have there all are always tautologies. You think the right thing to do is the thing you can do now but it's not true. And so we, if we persevere in further discussions tomorrow. One of the things we really should be focusing on is what, what we wish we had in the way of, of better models what we wish we had in the way of additional kinds of data, and I agree. We need more data at a variety of different levels and I think that the, the crisper perturbations other kinds of perturbations are helpful, and we saw that with Francesca's presentations, and it was striking how similar the results of Michelle studies to I think the ways that we've been seeing genetic architecture come out for some human quantitative traits so I think Naomi had had it exactly right that there will be parallels across systems in the way genetic architectures work. And can we do a better job of harnessing that information to help us with some of our questions and genetic architecture. And with that, I mean we really need those drink tickets, so. So there is a reception five to seven in the hotel in the democracy room of the hotel so if you're staying at the hotel you received to drink tickets that you can use at the, at the bar if you're not staying at the hotel and your local you're welcome to join us, and as the NBC suites democracy below our just up the street. That's a short walk if you want to walk back to the hotel you can walk or you can take the shuttle back to the hotel, which will leave now. Also, if you're checking out tomorrow you can bring your bag with you here, we can store it here near the kitchen. All right, so see you all in a bit.