 Well, thank you for that. And now we have a slightly different approach to the general question of application of multiomics to observational studies. You've heard four excellent talks so far with a little bit of discussion there. And what we're now going to do is challenge four panelists to within a very short period of time, five minutes each, to actually answer two questions or address two questions. Where do we want to be and what's the aspirational goal that we want to achieve? And what are the barriers and opportunities for achieving those goals? I will point out I'm Jonathan Haynes. I'm the chair of population quantitative health sciences at Case Western Reserve University. I've been studying a number of neurological and ophthalmological diseases for far too long. So with that, just remind the speakers they had the option of having some slides. So some of them will have slides, some may not. I will remind them that I will break in with the one minute point just to keep things on time. The first individual, the first person to take up this challenge is Dr. Miriam Furnage, who is a professor at the Center of Human Genetics and the Lawrence and Johanna Favreau distinguished professor in cardiology at the University of Texas Health Science Center in Houston. Her research interests lie in the molecular genetics of complex diseases with an emphasis on using functional genomic and genetic epidemiology strategies. Her focus is on the interaction between genes and environmental factors, and you've got five minutes to enlighten us on the challenges. Thank you, Jonathan. And thank you for the organizers for letting me give my perspective on multiomics. Let's see if I can get that. Here we go. So the first question we were asked was where we wanted to be. So I think I'm going to reiterate a few themes that we've heard over the past couple of days. So the first thing I will say we want a comprehensive set of omics information at a large scale. And I think that scale is going to be dependent on the question and the study design. Of course, we want a large number of people. We want multiple measures on multiple time during the lifespan, perhaps, and also perhaps different sources from single cells to various tissues. With these data, we hope to gain a comprehensive set of standardized and validated biomarkers of disease. I think Nancy started the workshop yesterday asking why there weren't very many biomarkers, so more biomarkers used in clinical settings. And I perhaps would offer that maybe we don't have as many validated biomarkers as we should. And I want to point some of the work we're doing as part of a Mark V CID consortium where we're trying to validate and standardize biomarkers for vascular contribution of cognitive impairment and dementia. So there is some spin-staking work that go into validating biomarkers that consortium is really doing some work there. With that, we definitely want to get to integrated models for disease risk and prediction for disease diagnosis and for therapeutic response. And so hopefully these integrated model would have an established clinical use in diverse population. And finally, the goal is really to have a routine application and adoption of these models in clinical settings and with integration of clinical tools and EHR. So that would be the aspiration in my view. Some needs and challenges and barrier. So I think each individual omics present its own set of challenges. And of course, multi-omics will inherit all of these and face additional challenges of harmonization and integration of heterogeneous and high-dimensional data. So I think we do a good job of identifying confounders on one sort of variation and biases across single omics. And so I think we have to carry that to multiple layers of data. One of the challenges as well is to have interoperable omic data resources and ontologies, sometimes annotation of molecule across platform within an omic, much less across different layers of omics is sometimes difficult. So having standardization there would be helpful. We've heard a little bit about advanced computational methods that account and for and leverage the complex interrelationship between and within omics. I think those are crucial. For that we need suitable computing infrastructure for data storage and management. And of course, these advanced computational methods such as machine learning, deep learning which require specialized expertise and training and therefore multi-disciplinary team science is definitely the needed here. The next step of a challenge I thought was also the need for accessing this information at large scale. And so I think that requires consortia-wide effort for data harmonization and curation. Not only shared data, but perhaps shared workflow and also computing infrastructure. So I think, for example, in top med, in charge, we've been using computing infrastructure that, for example, the commons, the bio-data catalyst that those are the type of infrastructure that definitely will be needed at large scale. And finally, in that same realm, public portals, I think we want to engage larger communities of researcher and clinicians and stakeholder. And so those are definitely something that we should think of. One minute. So additional needs, the need for tissues, disease-specific data set and the need to translate omics knowledge into clinically effective models for prediction. And that's for all. So we're going to need standard and best practice in developing and reporting these risk models and validated, again, validated model across multiple populations. And ideally, these models would be integrated into clinical tools. In terms of opportunities, I think we all recognize that multi-omics are more than the sum of each omics. And so we need to integrate these multi-omics to enhance our ability to understand causal relationship. And so we have genetics, it's very helpful in this setting, using Mendelian organization and also layering XQTL mapping, for example, layering omics into GWAS to understand causal genes. I think we understand also that the value of multi-omics for disease increases when it's integrated in environmental, social and lifestyle exposure, the exposome over the lifespan. And so I think we have to have an effort to really having this integration. And so I think that's where the longitudinal epidemiological cohorts that are exposure driven in their data collection can be very helpful. And so finally, we have a decade of GWAS studies given us the value of open science and data sharing and consortia-based collaboration. I think there is very good model for those. I think we should continue on this realm. And again, getting back to using more effective and meaningful predictive model in clinical setting, I think we can build on what we were starting to do in polygenic risk score, but perhaps supplement them with various panels of validated biomarker that reflect the lifestyle lifespan or the life stage and the disease stage and exposures, for example. Certainly for these models, we're going to have to take advantage of computational methods, advanced such as machine learning, deep learning, and above all, we need to keep the focus on addressing health disparities and improving on our diversity focus. So I'll stop here and let my colleagues continue. Thank you for that. And to remind folks, we are going to have the four panelists each talk, and then we will have about 25 minutes or so to have questions. So if you have questions that come along, you can put them in the chat and then we'll have the open discussion after that. So it's my pleasure to introduce the next panelist. Dr. Adam Buddenworth is a reader in molecular epidemiology in the Department of Public Health and Primary Care at the University of Cambridge, where his interests are in molecular epidemiology and revolve around the identification of genetic variation linked with coronary disease and related phenotypes in molecular omics such as plasma proteomics and metabolomics. Five minutes, off you go. Thanks very much, Jonathan. And thanks to the organizers for the invitation to join this wonderful meeting. It's attracted me on a Friday night here in the rainy UK. So thank you for listening. So why am I here? Well, my group spent much of the past 10 years working on a study called Interval, which is a cohort of 50,000 largely healthy people from the UK over which we've really tried to elucidate the path from DNA to disease. So anchored in genomic information and with linkages to e-health records on disease status. We've tried a variety of different platforms that you see here in blood and plasma based measurements, particularly around the proteomics and the metabolomics, really trying to understand pathway from DNA to disease. More recently in the last couple of years, we've kind of nationalized this effort by pulling together a 13 now population cohorts that again have genomic data, health record data and some form of multi-omics data with the idea of scaling and understanding reproducibility of cross cohorts that's been mentioned several times. Because these are population based, we can look at largely at things like prediction of incident disease outcomes, but also trying to understand etiology and causal pathways to disease rather than, for example, stratification of patients. So inspired by Eric Green's bold predictions for genomics, I thought the way to present my vision of what we might see in the next decade or so is the bold prediction for multi-omics and cohorts. And I won't go through all of this in detail, but just to pick up a few of the points, diversity of course been mentioned, I think there's a role here for both patient and population cohorts. I've had several people talk about the idea of serial measurements rather than what we mostly have in populations, which is cross-sectional measurements. I think multi-omic layers will be important. We've seen the more layers you have, the more information one can glean. And we're largely really only touching the tip of the iceberg with blood and plasma. So how do we get into, for example, different cell types, tissues, et cetera. As has just been mentioned, I think making the data widely accessible is a key point here and be used for the field. Of course, we need to make sure it's safely stored and managed. And of course, we're going to, as this data continues to grow, we're really going to require novel ways of thinking about this so we can really understand the complexity here. Let me show you a few examples of why I think this is important. My group doesn't focus so much on aging, but Benoit Lalier and Tony Wiskare at Stanford are interested in, because we made a data from the interval study open to people, they applied and accessed our data. And we're interested in looking at aging and the proteome and identifying peaks of aging and proteins that go up and down. One example I see here on the right is MMP12, which is a plasma protein level seen to go up in the peaks that they identified at 60 and 78. So it seems to increase as we age. And others had looked at this and shown that MMP12 levels seem to be positively associated with both primary and recurrent cardiovascular events. So the higher your levels, the higher your risk of heart attack or similar. But when we use genetic data and anchored that and genomic information, use the interval study, we identified that people who carry alleles that give them higher plasma MMP12 levels actually seem to have lower risk. So we think this is actually protective biomarker. It's released in response to myocardial damage. And therefore that gives you quite a different slant on the relationship between biomarker and disease. I think there is a causal effect, but it's potentially in the opposite direction to what you would see with just a longitudinal study. So we think longitudinal studies are important to capture that information before disease onset, but also anchoring in genomics for causality is useful too. Multiple platforms, I think maybe with genomics and then perhaps transcriptomics and some of the earlier stage technologies, we're more standardized, more advanced, I think with proteomics and metabolomics. One thing we've learned is that findings are often comparable across proteins. This is looking at PQTLs for two different plasma protein platforms we used. There is a good correlation, but not all the case. And you see kind of the horizontal and the vertical here, there are some exceptions. For example, this cis PQTL for a protein GDF15, which has a clear and striking signal in one platform, but a P of 0.5 in the other platform, completely different signals. I think we're only just starting to really understand the differences. And so using multiple platforms is going to be important. Finally, to some barriers, clearly cost of this genome sequence has come down dramatically, but not yet for proteins and metabolites in the same way. So I think coordinated efforts, these are scalable, making sure these are scalable and driving the cost down. Much of our omics were funded by industry, so I think partnering there is going to be important. I've mentioned already going beyond blood, we're just getting this snapshot with peripheral blood. Often the action is somewhere else, and we don't know how what's going on in blood relates to the tissue we're interested in. So that's clearly going to be important. Everybody argues we need more of these things. One of the challenges is where to prioritize. Is it more samples, more time points, more layers? And finally, how do we ensure we get that diversity and diversity at scale and diversity that's not a token mistake or sufficiently powered to tell us? So I'll leave you just with my bold prediction. Thank you. Thank you. That's very, very informative. Our next panelist to take up the challenge is Dr. Greg Gibson. He's the patent distinguished professor in the School of Biological Sciences and Director of the Center for Integrative Genomics and a member of the Pettit Institute for Bioengineering and Biosciences at Georgia Institute of Technology. His research interests lie in the use of transcriptomics for personalized medicine, predictive health, and predictive health. And Greg, you're off. Thank you very much, Jonathan. And thank you to the organizers. It's so wonderful to be here. It's been a great workshop. I want to make four points. I think whether or not there are opportunities or challenges depends on your perspective. I think they're a bit of both. So the first one is that I think we should have much more focus on therapeutic outcomes. I mean, predictive health is terrific, but from a patient perspective, using omics to understand how they're going to respond and how the disease is going to progress is important. So here's just an example. If you take on the y-axis, a sort of a classical histological measure, a Crohn's activity index, on the x-axis is a transcriptomic signature of disease progression and Crohn's disease. Either of them alone are reasonably good, but taken together. They give you much better prediction. In this case, in the bottom left quadrant here, we see almost nobody who's going to progress to the disease, such as the blue dots. So these different types of measures are somewhat orthogonal, but working them together, they can give you much better evidence about outcomes. The other thing about this is that therapeutic outcomes are usually 20 or 30 or 40, 50% of the population, which is going to give you much better precision or positive predictive value than if you're only looking at something like disease incidents, whether it's only 2% or 3% or 5%. So focusing resources on reducing things like the number needed to treat is a clear opportunity. The second point I wanted to make is that omic measures are a lot closer to disease and genotypes themselves. So if you take a typical EQTL, for example, where you have maybe a protective allele from GWAS and a susceptible genotype combination from GWAS, if you're doing genetic risk scores, you're based on just the genotypes. But if you think that actually what the genotype is doing is regulating gene expression, then it turns out that most people with the risk genotype have perfectly normal gene expression. So an orthogonal thing to do is to say, well, let's look at the risk based on transcriptomes. And if you do that, it actually turns out that you can get much, much better predictors. And on the right is an example, which we've got under review of predicting ulcerative colitis from rectal gene expression, where we see really, really good discrimination of collective cases, almost 50% prediction by transcriptomics, which you'd never, ever get with just genotypes alone, although we can actually generate a genotypic score from the EQTL that regulate those genes as well. The third thing running through quickly that I wanted to sort of say is we need to remember that all risk assessments are very much context dependent. So here are examples from the UK Biobank where we've just split the Biobank into two halves based on healthy versus unhealthy environment, whether it's obesity, hard CAD, type 2 diabetes, IBD. In every single case, the environment has a dramatic impact on your risk assessments. And I think this will be no different for audiomic assessments than it is for genetic risk scores. So in a typical example, you're seeing, for example, that the, so this is a prevalence risk plot where your risk, your prevalence goes up as your percentile of the risk score goes up. But in these, in the different environments, typically somebody with, in the 25th percentile and a bad environment has the same risk as somebody in the 75th percentile in the good environment. And those sorts of environments can be actual physical environments, they can be socioeconomic status, and of course, they can be ancestry based differences. So we've really got to be cognizant of the impact of not just ancestry, but all sorts of inputs we've heard many, many times this week, over the last two days on disparities in health. When we do our assessments, we just, we just need to take into account that those assessments have to be informed by our environmental background. And then the final point I wanted to make is that I, I think that actually single cell genomics as, as a great deal to us. So this is some data gathered with Judy Cho and Zubrik Agathasan and others in the inflammatory bowel disease consortium. And we can do our single cell profiling, in this case, of the epithelium of the ileum of patients with Crohn's disease. And when we actually go in and we look at the genes which are both GYs associated with disease, so they're probably causal and they're differentially expressed between patients and healthy controls. We actually see by doing a single cell transcriptomics, which cell types are involved. And so they're not all cell types. So for example, IBD is not a B cell disease, but we do see involvement of dendritic cells and inflammatory immunoregistry T cells. And we actually also see some engagement of epithelial cells. That's great. We, we may be able to get there as Tuli was arguing yesterday from deconvolution, but what you probably won't be able to get there is, is looking at patient specific responses. And I don't have it on this slide, but it turns out that if we go in, we can actually look at individual patients and see that either the cell type or genes within particular cell types are perturbed in very much patient specific manners. And I think the only single cell profiling is about the only way that we're going to get to that level of resolution, which you can then turn, take that a step further and say, well, this is the source of your pathologies. And this is a sort of therapeutic regime that you need to be able to get, whether it's directing at a goblet cells or an immune cells or whatever. And Mike Snide had a great example of that and diabetes yesterday from his own profiling. So if I was to sort of make some broad suggestions, it would be, first of all, that I think we, as many have said, we need more support for consortium based multiomic data acquisition, preferably longitudinally and preferably multi tissue. I really think that the integration of multiomics must be done in the context of the patient's environment, which includes socioeconomic, it includes diet includes microbiome, and includes ethnicity and ancestry. And I think that we need to pay much more attention to precision over accuracy. So forecasting patient outcomes. So accuracy is sensitive and specificity precision is positive predictive value. And I'd love to see in the next few years, we're using multiomics to improve the predictive value of our genetic evaluations. All right, thank you very much. And the last but not least panelist to take up the challenge is Dr. Allison Goat, who is the Professor of Neuroscience and Director of the Ronald Loeb Center for Alzheimer's Disease at the Icon School of Medicine at Mount Sinai in New York. Her research focuses on dementia and addiction with the aim of understanding the molecular basis of disease and ultimately identifying novel targets for therapeutic development. And she, of course, uses multiomics to get make progress in that area. So Allison, take it away. Hopefully you can see the right set of slides. Okay. Yes. So as Jonathan said, my focus is really on neurologic and psychiatric diseases. And so my first point really is around access to the relevant, relevant tissues for this group of diseases. And you can see here, this is looking at SNP heritability across a range of neurologic and psychiatric diseases, taking different cell types from within the brain. And I think sort of pretty stark contrast actually for Alzheimer's disease, you can see it says really in microglial enhances the SNP heritability, whereas across a range of, oh, sorry about that, of psychiatric and behavioral disorders, there's broader evidence of heritability across different cell types. And the reason this I think is really important is microglia represent a very small percentage of total brain cell types. And therefore bulk data from the brain doesn't get you anywhere at all in terms of being able to understand the disease. And so single cell data sets, I would argue, is something that are particularly important in this context. Importantly, we and others have found that there's a lot of overlap between the enhances in microglia and in peripheral myeloid cells. And indeed, when you sort of use peripheral myeloid data, you can actually help to understand a lot of the genetics of Alzheimer's disease, but not all. And I think that that's really the important thing, that maybe there's 50% of the loci from these GWAS studies that we cannot assign genes to the locus. And I think it's very likely that there are the specific differences in this microglial population that we don't pick up in peripheral macrophages or or in monocytes. And so one of our, I think, problems is that we really it's even hard to get to gene discovery because of lack of data for the for the cell types that are most important for understanding the disease. And so if we want to understand molecular mechanisms there, we need to have much better quality data, even from the right tissue and the right cell types, and across diverse populations in order for us to be able to understand this. And the same is going to be true. So we need this information if we want to be able to develop novel therapeutics or diagnostics and biomarkers. And right now, certainly in Alzheimer's disease, we have very few biomarkers. Up until very recently, the ones in cerebral spinal fluid have been the ones that have largely been used. I say there is promise there now in terms of plasma biomarkers. And so certainly anything that we can do with peripheral tissue that will help us will make things a lot easier than if we need brain tissue, because clearly this is not an organ that's accessible to us until people die. And that that's sort of significant problem in trying to disentangle cause from effect in these things. And that's really right now most of our omics data does come from bulk brain for obtained a autopsy. And I think that's a significant limitation. Right. So I think that we need certainly we need more single cell data, but we need a purified cell cell data. And we need a lot more different types of data in order to understand this underlying genetic risk. In terms of opportunities, I think so thinking about longitudinal studies, I mean, also all the sorts of omics data that people have mentioned, I do think, you know, obviously we we will need CSF and brain, but otherwise blood and omics from blood is obviously the most accessible and clearly going to be broadly useful. I think you need longitudinal studies, you need good and deep phenotype data, and genetic data to identify novel biomarkers. The examples I've given here specifically to AD, but I want to illustrate particular things. So the first two examples are longitudinal observational studies in fat families with known causes of disease. And I think we shouldn't forget that there is some value in being able to look at these because we know who's going to get disease, we know when they're going to get disease. And so we can do longitudinal studies in these people and and they're very motivated to participate because they're in high risk families. And so there's a lot of value for those kinds of families. We also have, you know, sort of late onset families as well, where there are large cohorts. And then there are case control data sets, some with longitudinal data, and where we could collect omics data. And then the last example is obviously biobanks, where I think there's also great potential. And the idea of using polygenic risk scores within large biobanks to be able to identify at risk individuals and to be able to collect large scale omics data longitudinally in these kinds of samples. And in addition, I would add in addition to these omics samples, I would argue that that we should be banking IPS lines or at least PBMCs from patients so that we have lines that we can differentiate into any cell type that may be useful for the disease that we're interested in, that we can do omics on those cell types. And we can study function and use them also if we have diverse panels, use them for drug screening as well. So I think there's some great value in having not only omics from fluids, but collecting cells from people as well. So I'll stop there. Thanks. All right, thank you very much. Thank you to all four of the panelists. I would ask that the panelists come on and come on the video so that you have an opportunity to everyone can have the opportunity to see you and ask questions. So we have approximately 20 minutes for questions for our panelists. So anyone who has a question, raise your hand right in the chat or speak up. So, Eric. Thanks. Alison, that was great. I would enthusiastically agree with the first half about the relevant cell types and tissues needing more data. But I guess I worry that we're rather naïve that we don't know what the relevant cell types and tissues are for most late onset chronic diseases. We think we know some of them. Well, first of all, I don't think there's one. And second, we think we know some of them. And I worry sometimes the cell types and tissues that we point to are the cell types and tissues of the symptoms, not so much the causality. I mean, I would agree with you that I think for many diseases we don't know where most of the SNP heritability is because we've been using bulk tissues. And if any of the risk is in cell types that represent minor cell types there, and we're probably missing it all together. But I think it's a really important component of trying to understand the diseases to know the contribution of different cell types, at least from my own perspective. I had a very neuro or neuronal centric view of Alzheimer's disease until about six or seven years ago. And now maybe I'm proselytizing like they're newly converted. But it's sort of, I think it really opened my eyes to how biased we can be about thinking about a disease if we're thinking about it based on symptomatology without really thinking about the causal role of particular cells. So I think it's really an argument for single cell data, right, that with bulk data it's extremely difficult to get beyond the most common cell type in that tissue. Yeah, thank you. Okay, Judy, I think you're next. Okay, I had two questions, one for Greg and one for Allison. For Greg, it was interesting that diet seemed to have the least effect of inflammatory bowel disease. Did you look at Crohn's versus Ulcerative Colitis and kind of how scalable is that? And for Allison, kind of smart watches or devices plus Alzheimer's? Yeah, well, I know who wants to get first? I can if you like. I do think that passive data collection in an Alzheimer population is going to be better than an active one, right, in the sense that, you know, as you, as the disease develops, you're not going to be actively entering data, but definitely wearables where you can collect data I think can be very valuable, right? I mean, I think sort of loss, sort of subtle behavioral changes can be picked up with wearables. And even collecting memory data more frequently has been shown to be much more valuable than collecting memory or more accurate than collecting memory data once a year when someone feels really stressed because they're coming in for their test. Even people who are, you know, in the non-demented but at risk feel very stressed by doing a memory test and anything could make that poor. Whereas if you're collecting data, you know, on a smartphone or something every week or every few months, a couple of months or something, you're going to get a much better picture actually of what the real state of someone's memory is. So, yeah, definitely I think that there's a lot of value in having wearables and for data collection. Or just smart data collection generally, right, rather than just a clinical interview once a year. So, so, Judy, I'll send you the paper, the preprint we put on BioArchive in the next couple of weeks. But yes, there are differences between UC and CD in the environmental impacts on the polygenic risk assessment. You know, nicotine and alcohol consumption are big factors that really discriminate individuals. But it gets to this idea of conditionality. Actually, you know, dried fruit and fresh fruit consumption produce different results. But I don't, I doubt that fruit consumption itself is the important thing. I think their mark is more generically of the type of diet one has. The other component that comes into all of these factors is actual familial disease, presence of disease as well. I think that's probably partly genetic, but I think it's also other aspects of the shared familial environment. But actually, having a shared relative with a disease has a major, major impact on your relationship between prevalence and polygenic risk in ways that we didn't necessarily expect all along. And I was going to actually add to the discussion before your question. It's not just finding the right cell type in my view, but it's finding it in the right circumstance. So a B cell or a T cell, you know, in the gut, it's not the same as a B cell or a T cell in the lung or a B cell on a T cell circulating in the blood. And it's certainly not the same as one at diagnosis as somebody in with established disease. So I think we need to be doing that sort of profiling, not just longitudinally in sense of following patients, but at different time points in different conditions so that we actually get a full spectrum of what the environmental impacts are on the profiles that we generate. Thanks. Okay, Tiago. Thank you very much. So I'd like to bring up a topic that I think is really important. And I haven't seen somebody, you know, tackle this directly. And so this was alluded to yesterday by, you know, Nancy Cox and to the La Palana, where they basically, I think Nancy was more focusing on the importance of multi-omics to identify new biomarkers, and then Tuli was very much talking about the importance of single cell resolution, cell specificity, et cetera. And I think that it might be important to discuss the considerations that need to be taken into account for, you know, for example, what study design and also technologies used based on, you know, in relation to what your goal actually is. Is your goal to understand disease etiology from a really sort of mechanistic perspective, or is it to identify biomarkers? Because I think that it's particularly in a clinical setting, I think it's really important to think what kind of technologies and analyses and computational tools we need in order to address these. So I was wondering if the panel could discuss this a little bit, basically, how can we use multi-omics to identify biomarkers versus disease mechanism? And are these incompatible or not? So maybe I can start. I mean, my observation I'd make, because I don't think they're really separable. So when we did our transcriptional risk score study on IBD a few years ago, a part of it that really struck us was what we call difference between coherent and incoherent associations, which I think was really getting a mechanism. So what this is, is that you'd expect an EQTL, where the risk a little increased expression, you'd expect to see increased expression in cases relative to controls. But very often you didn't see that. In fact, it was the opposite way around. And that made us scratch our heads. But our interpretation of it is that in fact, what this combined data is telling us is that in some cases that elevated expression is actually protecting the individual. And if you have the allele that sends you in the opposite direction, then you can't protect yourself enough. So it's telling us that GWAS associations sometimes are not promoting disease, but they're part of the protective response. So I think in that combined effort of trying to get biomarkers that were predicting disease progression, we were getting insight into the mechanism of what was going on. And so that's part of my response is that I think that I agree with you that different algorithms are going to give different answers for different types of questions, but they're not separable processes. I'd like just to add very quickly, because my point was, for example, I see a lot of studies, multiomic studies that, for example, use whole blood traditionally, so tissues that are easy to access, but then they're studying, let's say, neuropsychiatric diseases. And it's not always clear in multiomic studies that people are using the tissue that is the most relevant for the particular disease they're studying. And sometimes frustratingly, these studies don't find anything, but I would want ideally to hear from you whether should we still be using whole blood? Should we just go full on on single cell analyses? What do you think? Well, I guess Tiago, one response to that is the recent Alzheimer data where plasma levels of phosphotide tau look like they're as sensitive, if not more sensitive than measuring the same protein in cerebral spinal fluid or imaging-related proteins. So I think that that would argue that while you might not intuitively think it's the best place to be looking for a marker of a neurological psychiatric disease, it can actually come up with things that are very powerful. I would agree with that. And I think we've found signals in blood for, you know, particularly when we think about proteins where maybe the single cell and the cell type specific hasn't necessarily come as far, or at least in as large sample sizes that one can pick up signals that you wouldn't expect to see maybe in blood, but it points you towards an etiological pathway in disease that you can then follow in tissue. I think to your first point, Tiago, I completely agree. And I think we've heard during the last two days that it really depends on the type of question you want to ask. And if you're looking at response to treatments, then you're maybe going to take a patient cohort we heard from Mike Schneider about, you know, thinking about this idea about people's veering off their trajectory, moving away from wellness into kind of health conditions. I think if you want to then think about, you know, predicting disease, then you want kind of incident, kind of longitudinal cohort. So I absolutely agree. We're going to, you know, there's no single answer here. It depends on the question you want. And design is crucial to that. Yeah. Okay. We'll move on. Phil, you had a question. Thanks, Jonathan. So let me have a question for Miriam and Alison and for the other powers, but maybe using Alzheimer's and the brain as an example, you know, to what extent do you think that we can leverage the existing studies, you know, to generate this type of multiomic data versus designing the proper study that we may want to do today based on the many things we've learned in the past decade or two. When a lot of the cohort studies were already going on and, you know, we're going to design to do other things. And in particular, I'd say the big divide is the separation between the end organ sort of data that we get from autopsy material, in this case, the brain versus the fluid biomarker, the fluid data that we can get from CSF and blood, which are on different subjects, typically much younger than the individuals from whom we get the brains. Yeah, that's a difficult question, Phil. So I think there may be some advantages in using the cohorts that we have for convenience, certainly having multiple omics on the same person is probably going to be advantageous versus cobbling together omics from participants, a set of participants that may or may not be similar to another set of participants on which you'll have a different set of omics. So I think that's that's a major, that would be a major advantage is having to comprehensive set of omics on a decent amount of participants, all of them. And, and obviously that set of participants would be selected based on the question you're asking, right. So for some question, it may be that longitudinal cohort may be appropriate for some others, maybe that the data, the samples that we have on Alzheimer's disease patients may be more appropriate. But that's a different, that's a very difficult question to address. If we have to do something de novo, I think we better think clearly in our study design and the question you want to ask to make those decision rationally. But for existing data, I think there is also a merit of putting together existing data that's already been collected. But again, I think that's the harmonization is going to be very difficult. But you know, that's a zero cost, it's already been collected. So I think there is some merit as well into looking into trying to harmonize this type of data to ask reasonable questions. Yeah, I don't think I really have much to add. I mean, I agree in both ways that I think that that ideally I think you want the different omics data sets on the same individual and that may necessitate us as you know, collecting either new data on existing cohorts or creating new cohorts. But I mean, I think just from a very practical point of view, some investment does need to be made in trying to harmonize and optimize what we currently have as well because we will learn some things from that. And maybe that will lead us to design better studies for the next generation of studies. But we definitely need these kinds of data on the same individuals. And I do think one of the challenges in Alzheimer's disease and what we have so far is that it's mainly from dead people. Right. And yeah, we just we said earlier about how it's probably got a 20 year course of disease. So I think that that that's a real challenge in this field that we really need some data from existing longitudinal cohorts where we can sample people at different points in the disease. And that's what I tried to illustrate actually at the end of my talk with suggestion of some of these, you know, some of the certainly rare forms of the disease, but predictable in terms of when they're going to onset disease and who's going to get disease and collecting omics in those kinds of families might be extremely valuable. Okay, we're running up on the on the hour. I'm going to give Zhihang the chance to ask her ask the last question. And then I think we're due for a little bit of a break if I'm if I remember correctly. So Zhihang. Thanks so much, Anson. Great discussions. My questions is about how to effectively design multi-omic study in longitudinal studies. In particular, ideally, we want to sample everybody and in the whole cohort think about like a million people and get the multi-omic data. For example, yearly multi-omic data for everybody. But in reality, this is not possible. And so we need to sample a subset of people. Then how can we effectively sample a subset of people in multi-omic study to help with the scientific goal we are interested in as a panelist that discussed. And so the traditional way is in genomic study is to do random sample. But in epidemiological study, we know this is not effective way. And it's not a powerful way to use the limited funding to do. So for example, in epidemiological study, when we'll use like case cohort design. And so suppose you have for each new case, and then for those people at risk, at risk cut, you random sample a subgroup of people. And so this will improve the power because you will sample enough cases as time moves on. And also you have sufficient controls. So by this type of efficient epidemiological design has not been used much in genetic genomic studies. I want to get the south from the panelists how can we encourage incorporating those more efficient epidemiological designs in designing the multi-omic study in longitudinal settings. And that's a key challenge. You hit on there and there's different approaches. One is to take the UK Biobank approach, which is to say everyone's got their own pet disease of interest. So you either measure something in everybody or in a random sample. So you don't favor any particular disease. But of course that has limitations. I think the rationale for doing that is to avoid the piecemeal approach whereby different parts of your cohort have different multi-omics data because everybody's interested in a different disease. And then you end up with this sort of patchwork that you can't knit together. But I think if you want factors to look at, I think PRS as somebody mentioned earlier, I think there's a lot of interest now in identifying either related family members or people with high polydent risk in the age, as people said, for Alzheimer's as well, hitting the age bracket that you want. So I think there are ways to enrich. But I think if you're thinking about a large population cohort or even a patient cohort, how you do that so you don't end up with a piecemeal jigsaw is can be quite challenging. Okay. I think we've unfortunately run out of time. I'm sure it would be more questions and more answers. We could talk about this for a long period of time. But we are due for a short break at this point, I think. So I'm going to turn it back over to Joanne L and Erin. And they can take it away. Thanks, Jonathan. Yes, we are due for a break right now. And everybody come back at 315. We would encourage you to keep connected to Zoom so that we don't have to let everybody back in. So go take a break and come back at 315. Thank you.