Scientist Stories: Karsten Suhre, Genetics Meets Metabolomics

series. Today's webinar from Dr. Carsten Zurich will focus on Genetics Meets Metabolomics: From Association to Translation. Before introducing our speaker, I want to quickly review just a few logistics with you. Should you experience technical difficulties, you can contact us through the questions box in the expanded control panel or by phone or email. If you need to view live closed captioning, please click on that link and it will appear in the chat box. At the end of Dr. Zurey's presentation, we will allow time for questions, so we're going to let Dr. Zurey present and then we'll have a question and answer period following his presentation. If at any time during the presentation you have a question, please do, though, type it into the question box, and then we will ask it during the Q&A session. So as we go through and you have a question for Dr. Zurre, go ahead and type it in. We'll have them in the queue, and then at the end when we have the question and answer period, we will begin to ask those questions. So it is my great pleasure to introduce Dr. Carsten Zurre. Dr. Zurre is a professor of physiology and biophysics at Weill Cornell Medicine and the director of the bioinformatics and virtual metabolomics course at its branch campus in Doha, Qatar. He holds a PhD in atmospheric chemistry and meteorology from the University of Toulouse in France and graduated in physics from the University of Osnabrück in Germany. So we welcome Dr. Zurey to the call and look forward to an exciting presentation.

All right, so I guess I'm on now. So, I'm very happy to have everybody online. So, it's a bit strange to talk just to the screen, but I'm happy to have you, and I will try to do my best to make it as exciting as possible. So, as Chris already said, I'm talking about genetics meets metabolomics. So, we'll try to explain how genetics and metabolomics work together, and then also in the second part of my talk, I will try to give a bit more of an explanation how things could translate to application. So, I don't have any conflicts of interest, and let's just go right away into the topic. So, what makes us different? I think the question is always why are some people getting diabetes while others don't, although Maybe the slim guy could be the one who can, who might be eating a lot of hamburgers while the other may be healthy, healthy eating, but he still gets diabetes. So, the question is what is this? And, of course, these are things like our lifestyle, whether we, what we eat, what we drink, our lifestyle, whether we smoke, whether we do sports. So, this here is the logo from the Qatar Sports Day. What drugs we take, of course, and then recently also has been shown that, of course, the gut microbiome may play a major role in our health. And, what would be the central topic of my talk today is, of course, the counterpart of all our environmental and lifestyle factors. This is our genes. All right, so, as I said, all these different factors make up whether we are predisposed more or less to complex disorders, and the question we have is really what makes us different. Alright, now to better understand what happens in our body, a very good way to go is to look all— at all the small metabolites, all the small molecules in our circulating blood, in our urine, saliva, any bloody fluid. So, I always like to see this as some kind of visualization technique or an imaging technique where we can actually visualize what's happening in our body. In our body. And metabolomics is actually the tool that really helps us do this. So, what you see here is one visualization of a metabolic pathway map where all the dots would be all the different small molecules that circulate in our body. You would have here, for instance, the glycolysis with the TCA cycle, and all our different metabolic capabilities that are encoded in our genome are basically connected together by the lines. So, the dots are all of our molecules, our small molecules, glucose, fatty acids, amino acids, and so on. And, the lines are basically the different enzymes, transporters, and everything that acts on our metabolism. So, just to come back a bit to the idea of the imaging technique, metabolomics, so it's really about metabolomics here, of course. Metabolomics has been a technique that has been around since the Middle Ages. You want. So, here, it would be a urine chart where already there would be multiple parameters to measure, like the taste, the smell, the color of the urine. Then, biochemistry came along to measure more precisely different biochemicals. This here is an ad from a mass spectrometer from the '30s. And, what really brought about the new omics era, the metabolomics, and, in parallel, also the proteomics era, are now these modern mass spec. Centimeters. And, if you parallel this with a number of metabolic parameters that can be determined by the different techniques, you very quickly see the parallel I mentioned between imaging and metabolomics. So, the more parameters we have, the better the techniques are, the clearer the picture becomes. So, now we come back to trying to understand disease. How does it come to complex disorders, especially complex disorders? Of course, there's one kind of association that has been started with the advent of the microarrays is the genome-wide association studies. And, well, I will talk a lot about genome-wide association studies here. So, genome-wide association studies have been done with a lot of endpoints, especially diabetes, kidney disease, cardiovascular risk for cancer, and so on. In parallel, big— epidemiological studies have run and have studied predisposition to disease. So, epidemiological studies have shown that smoking, for instance, can give you lung cancer, and that sports may be good to reduce your diabetes risk. However, the problem is that all these complex disorders have attached very small effect sizes, actually, if you go to the link between the causative factors and the endpoint. For different reasons, so it's pretty difficult to really find out what is the role of all these different factors in the etiology of the complex disorders. And, here comes now metabolomics as an intermediate phenotype. So, the idea is to say instead of measuring the direct link between genotype and endpoint or lifestyle and disease endpoint, we put what we call an intermediate phenotype in the middle. Like all the different small molecules in the metabolites. And then, we study association between metabolites and disease. So, these would be metabolome-wide association studies. Sorry, then you can also look at studies with a metabolomic endpoint. So, you can see how does smoking change your metabolome, for instance. And, the third kind of study would be a genome-wide association study between genetics and metabolomics. And, that's the part I would like to focus on today. So, the topics I want to cover in my talk are basically now two halves. So, the first half would be on human metabolic individuality. So, here I will be speaking about all the genetic variants and the association studies with metabolomics and what they teach us. And, in the second part, I will try to go beyond that and try to go beyond association to the translation How can this knowledge that we actually get from the GWAS be translated to application, like modifying our metabolism, working on drugs, precision medicine, and in the end, I will just give you a little bit of resources that you could look up on. All right. So, human metabolic individuality. So, with the era of genome-wide association studies, GWAS, came, of course, at some point the idea, can we not only connect GWAS with diabetes and cardiovascular and other endpoints, but also with intermediate phenotypes. And, I'm proud to say that our group in Munich at that time, we were the first to actually perform a GWAS with metabolomics, which was actually a pretty low-powered study at this point. But, however, even although the number of samples were very small in this study, it very clearly became evident that actually the effect of these genetic variants is pretty strong, so that you can find a lot more information out of GWAS with intermediate phenotypes than what you would get out of GWAS with clinical endpoints alone. So, this led to what we now term the genetically influenced metabotype. So, imagine that you have a human population here, and then you look at genetic variants, SNPs, which are Typically, I'm looking here at frequent variants. So, for instance, if you take the prototype of one of our associations, is a genetic variant in the FADS1 gene, so the fatty acid desaturase 1. In the population, you would basically have people who are majorly homozygotes of one allele. You would have the heterozygotes, and you would have heterozygotes— homozygotes, sorry, of the minor allele. So, there are basically 3 genotypes: homozygotes of the major allele, heterozygotes, minor homozygotes. And, what we found in these GWAS, and I would come back to that repeatedly after this, is a statistically very significant association between genotype and the metabolic readout. So, in this case, we would find a genetic variant that is actually located in or near the FADS1 gene, and it would associate with the ratio between the product and the educt of the FADS1 gene. And, that's really something which is extremely surprising, because if you start with this non-targeted hypothesis-free approach where you would say, I look at, in this case, 500,000 different genetic variants, SNPs, and test every one of these 500,000 against 500 different metabolites, the chances that you find a significant association exactly between a genetic variant that is in the enzyme that actually works on the metabolites that I find in my association is really extremely low. So, the p-values here, later on, with larger population studies, they go into levels that you can say this is absolutely without any doubt specifically highly relevant. But, beyond that, what's important to note here is the effect sizes. So, in this case, for instance, the difference between the green and the red genotype carriers would be basically a factor of 50%, so that we can say that these are fast, medium, or slow metabolizers, actually, for the FADS1 reaction, which produces arachidonic acid. Which, again, is a very important precursor of prostaglandins. So, you can already imagine that this genetic variant will have a certain impact also on health, and I will be coming to that in a moment. The other important concept here to think about is that this here is one, just one variant in one enzyme working on one set of metabolites in our body, and given that we have different genotypes here, I like to imagine these as different settings of a valve in your metabolic piping system, and then if you look into the population, people will have different settings of these valves. So, they will be 0, 1, or 2 positions to set these valves in high, slow, or medium throughput. And then, if there are other enzymes associated with other metabolites, and I will come to these also in a moment, then there would actually be a network of different settings that would make up what we then call our human metabolic individuality. So, everybody would be basically defined by a set of valves set in our body, and the flux of all the metabolites through our body and how the input from the outside and the effect of the life factors actually play out on the disease risk depends on the setting of these valves. Now, this is nothing new. Actually, for me it was new because I'm a physicist. I didn't know Archie Buitgeroit before we did this, but I think for every biochemist, he knows him from well, from university, I should say. So, Archibald Garrett is the person who invented the idea of inborn errors of metabolism. But, he not only did that, he actually wrote already in 1909 that inborn errors of metabolism, well, they are merely extreme examples of variation of chemical behavior, which is probably everywhere present in minor disease. So, he already predicted What we found in our GWAS that there are actually variants in our genome that change, like the FADS1, the throughput of the enzymes, but these are not monogenic disorders. These people are not sick. They just have different settings of their valves. And, furthermore, he also already predicted complex disorders. So, in 1909, he actually suggested that there is a link between what he called the chemical individuality. So, today, I rather call it metabolic individuality. And predisposition to disease. So, if you have the opportunity to read this book, it's a really great writing, and basically something that could have been written in our days. All right. So, now that we know that these variants are out there and how we can find them, the idea is now how can we really catch them. So, we have the roadmap now is to map out the atlas of human metabolic individuality, to identify ideally all relevant genetically influenced metabolotypes. So, we may discuss what is relevant. I would say at least the ones with large effect sizes and probably also with large allele frequencies. And then, translate this knowledge to clinical and biomedical application. And, that's what I will be talking at the end. All right. So, I think Krista mentioned it already. I was, before moving to Qatar, I was working for a long time at the Helmut Zentrum Munich, and what you see here is the metabolomics core facility at the Helmholtz Center in Munich with the mass spectrometer in the back, robotics, and sample prep here. And, what we did is, over the years, we ran a series of genome-wide association studies using different population studies with increasing number of people, and also different types of fluids like blood and urine, and recently also saliva. And ran them also on different metabolomics platforms because they are, of course, pretty, pretty complementary. Like, there would be Metabolon, which is a non-targeted, very broad approach. There would be BioCurtis, which is very targeted but quantitative and more lipidomics-oriented. And then, there are NMR techniques. So, we used NMR in urine with Kinomics annotation and Numaris for lipid class annotation. And, over the years, we ran different GWAS that I mentioned here. So, at present, as it stands, I think we're at about 150 or so genetically influenced metabolotypes identified through all these GWASes. And now, I will go and show some of our examples. So, one of our key papers was GWAS that we ran using the Metabolon platform. And, in that study, we used 1,800 samples from the KORA study, and then 1,000 samples from the Twins UK study. And, what you see here is the Manhattan plots. That is the p-value of association by chromosome position and lumped over all metabolites that we have measured. And, what we find is that there are 37 genetically influenced metabotypes that reach strong genetic, strong genome-wide significance, and that can also be replicated in the twins. So, these are 37 metabolites that are really very strongly confirmed, and we confirmed them later in bigger studies as well. So, now I will go a bit through what we learned from these. So, I think I mentioned that already before for FADS1, so the amazing thing, what I always really still surprises me today is that you take 2 drops of blood, you run one through a genotyping array, which is by its own highly technical. You take the other and run it through a mass spectrometer, and you treat— you do all the texts of the— all the steps of data analysis and QCing and so on, and at the end, you do the statistics, and you look for the variants that have very, very high significant p-values. And, what turns out is that you find, actually, a match between the gene where the genetic variant is located in and the metabolic phenotype, which actually matches in very, very many cases the function of the gene. Like, FADS1 is the delta-5 fatty acid desaturase, and it associates with the ratio of the product and the educt of FADS1. That's basically our prototype for all the genetically influenced metabotypes. But, there's a whole list of these, and I'm just showing some here. So, SCD, for instance, is another fatty acid desaturase, and it goes with its precursor and its substrate. It basically puts in the delta-9 double bond into the long-chain fatty acids. Then, you have things like proline dehydrogenase that also is with proline, and glutaminase with glutamine, and oxoprolase with oxoprolene. You find nucleotides, the 5' nucleotidase with inosines. And then, even, there's a little bit of proteomics here because you even find dipeptides with peptidyl dipeptidase, the ACE enzyme. I come back to that also in a moment. Or, acetyltransferases with acetyl amino acids. Right, so, given that this is really a very, very amazing match, the question is now, what can we learn from these things? And, I think even the fact that every association tells a story is, in a way, a translation of the knowledge in its own. So, we present here our 37 hits by metabotypes. So, associations of different variants with lipids. Here's FATS1. Polypeptides. I said there are also some small proteomics things, if you want. Xenobiotics. There are a lot of variants also in transporters, nucleotides, carbohydrates, and amino acids. Acids. And now, here would be, for instance, the FADS1. That's our prototype. So, let me just tell you a little bit more about that one now. FADS1, I said there is— it plays a major role somewhere in the fatty acid pathway. So, you would expect that people who have different valve settings in FADS1 would show such an effect. And remember, that study was 2008. This paper here is from 2010. So, this is basically a confirmation of what we would predict from our data. It's an interaction between the intake of omega-3 and omega-6 fatty acids, which are the substrates of FADS1, FADS1 genetic variation, and the outcome HDL cholesterol. So, you can actually use these genetic variants to predict the outcome of a certain dietary intake as a function of of genotype. Another example is pharmacogenomics. SLC01B1 associates with a ratio of different fatty acids, and that's a very well-documented pharmacogenomics locus where the genotype, the CC genotype, actually gives you a 20% chance of getting statin-induced myopathies, while the other two genotypes are nearly neutral. So, for this one, we could now, for instance, take the information we got from the GWAS with these fatty acids and say these fatty acids have probably to play a role in this phenotype and create hypotheses on that basis. Yeah, then functional genomics. Functional genomics is the question of how can we really understand the human genome, so what can we learn from it. SLC16A9 is a transporter that had only been coarsely annotated by homology. We found an association with carnitine, so you would create— you would hypothesize that this transporter probably transports carnitine, because if the carnitine levels change by genotype, this is what it should transport, right? So, we were lucky at that point that I had a good collaborator in Oxford, and he could do the experiments. So, we actually created the hypothesis and then collaborated. I said, can we do the test? And, what he did, he used a frog oocyte, isotope-labeled carnitine, and could actually show that, yes, SSC16A9 is transporting carnitine. It's actually an efflux transporter for carnitine. So, this is a way of doing functional genomics. Then, comes the creation, basically, of hypotheses for clinical studies. So, AKAT-S was also in our first paper in the Giger et al., 2008. So, what you read here is basically from the abstract of a paper I will just show you. So, they took up our discovery and they say this variant here could play a role in the fatty acid, in the beta-oxidation. And, see, this paper is from 2011. So, they actually did the test. They tested people with oral glucose load and they identified and confirmed that this genetic variant is actually associated with reduced insulin release following an oral glucose load. And, this kind of finding, you could never have done that in a non-targeted way because then the statistical power wouldn't be there. Here, you'd say, "I just look at it as an outcome on insulin based on this one genotype, and then I only need a small number to prove my case." So, one other application of this is to create clinical studies. Next thing would be— Again, something like pharmacogenomics. So, I said the ACE enzyme, which is the target of the ACE inhibitors for blood pressure, if we go into the KORA study, so to the left here you have the KORA study, we would have people who actually take ACE inhibitors. They, these are the levels of the dipeptide aspartylphenylalanine for the people who take the ACE inhibitor, and the ones who don't take them, they show a very strong genetic dependence. That's the genotype we observed., and the ones who take the inhibitor all have very low levels. And, the same thing is replicated in the twins. So, what this basically means is the difference between the treated and the untreated is a function of genotype, which suggests that there could be, or should be, a very strong genotype-dependent effect, which you could then, if you look deeper into what the ACE enzyme really does, suggest things of, like, for instance, genotype-dependent side effects from the ACE inhibitor. So, I don't go into the details here. I can— you can look at that later. There are actually a lot of other dipeptides, so we can actually make a hypothesis about how these genotype-dependent side effects of ACE may work. Then, a third, or another interesting case is we have different genetic variants that actually act on the same metabolite. Or, actually, in this case, I should even say it's more than a metabolite. It's a ratio between a polypeptide and a phosphorylated polypeptide. So, metabolomics also captures that. So, this would be a measure for fibrogen A-alpha phosphorylation. And, there are 3 enzymes that associate with it. So, here you see basically additive effect of the 3 genes with a total of 6 risk variants and replicated in the twins. And, what you can create from here now is a model of basically how these things interact. So, ABO, we have shown that ABO is associated with fibrin A-alpha phosphorylation. That phosphorylation is associated with blood coagulation in other studies. APL, we also have the association here, and FAUT2 as well. So, basically, all the three associations we have, they go with blood coagulation, and then you could put in knowledge that actually supports this. APL is a phosphatase, so we talk about phosphorylation here. ABO is in another GWAS associated with ALPL levels and with venous thromboembolism. So, again, you come here to blood coagulation. Then, actually, FUT is coding the H antigen that is actually glycolyzed by ABO. And, it's the precursor of the H antigen of the blood group. So, the whole thing really makes out a total story, and together you would now speculate that there's a cumulative additive effect between these three genotypes on cardiovascular and venous thromboembolism risk, because that's what these variants are associated with. All right, and bring this now to the disease endpoint. So, if I take all these variants here I have, and I overlay them with GWAS hits, so this was in 2011. They are more like that. So, the red ones would be disease risk loci, and the yellow one would be intermediate risks like HDL or LDL to disease, and the green ones would be other GWASes like body height. So, what we find here is that there are actually many GWAS hits with complex endpoints that are also hits with metabolomics. And, that brings me back to my initial statement that there are actually intermediate phenotypes between the genetics, the environment, and the endpoint. And, in our study, I could now replace this. So, remember the ones I've showed you, FADS1, here the fibrillin A alpha. So, again, if you build up a network of where every arrow here is an association to the endpoint of coronary artery disease. All right, so, to sum human metabolic individuality up. So, we find a match between metabolic and gene function, surprising large effect sizes, frequent genotypes, that's important. Overlap with GWAS loci, intermediate phenotype to disease, gene-gene and gene-environment interaction. I've shown you the examples. Pharmacogenomics and functional genomics. And, with this, of course, I think you see the motivation to push things further. So, we were lucky to get funding and support from Pfizer. So, together with the Twins UK, we then went on and did a study on creating what we now call the atlas of genetic influence on human blood metabolites. So, what we did here was we used nearly 8,000 people from the Twins and KORA, 2 million SNPs, 500 metabolites, a hugely dense Manhattan plot came out. So, we found 145 validated replicated genetically influenced metabotypes. And now, I cannot put these in hexagons anymore. So, every Every name here is basically one association of a gene, and then you can classify them by their properties. And, one thing you see right away is there are a lot of enzymes, a lot of transporters, and only very, very few for which we actually don't know at the moment why we find the metabolic association. So, these are the ones like the SLC16A9 association where I would actually go and look for functional genomics and try to understand why this oscillation is there. While these already are very clearly metabolic enzyme, metabolic individuality loci, where every variant in each of these genes actually is one of the valve settings in your organism. And then, we can also annotate them. So, the ones here in these Venn diagrams are some that overlay already with drug targets, metabolizing enzymes, and transporters. Archibald Garrod says that there would be moderate variants of monogenic disorders. So, for many of these genes that we find, there's actually a corresponding monogenic disorder existing also. And then, the blue ones are all associations from the GWAS catalog or to drug response. So, these are all loci where the genetic variants would already be— the metabolites would already be an intermediate phenotype to the endpoint. So, I think there's a lot of stuff to be looked at. From this, you can even blow up and create more complex networks where you would put in— the circles would be the metabolites, and the diamonds would be the genetic loci, and the colors here show you different gene endpoints. So, this would be basically now a network of coronary artery disease where there would be here the genetic individuality., and also the metabolic readouts. And, here is the corresponding GWAS locus that go into this. And then, the next level would be to bring everything together. So, what you see here is, this is the, what we call the atlas. So, this is 145 genetic associations with nearly 500 metabolites. This is already a condensed network, so I, in the end, I will give you the link. The detailed network can only be viewed with specific tools. All right, so I think I have to be careful not to run too much out of time. So, now I come to the translation. So, the question is now how can we translate this knowledge into applications, and I show you quickly 3 different applications. So, one is what we call rational metabolic engineering. So, here the idea is, so we come back to my initial transparency, so what makes us different. Is, of course, gene, lifestyle, environment, gut bacteria, genomes, everything. We can see how that plays a role on disease, how that goes into the metabolic pathways, and then, basically, if you go from the metabolic pathways to these, to the atlas, so the atlas is actually now a measure of where all these valves are sitting. And, the way I like to see this, for every genetic variant here, depending on genotype, you have different throughputs of this different genetic loci. So, this would be one valve for FADS1, and in this network there would be 145 valves. So, if I look at this more from an engineering point of view, this would be the interior of your metabolic system in your body. So, there would be inflows and outflows and factors that play a role on that.. And, the valve settings here, that would actually be the ones I know. And, the idea now is to say every genetic association is basically an experiment conducted by nature on humans. So, nature experimented on us and tells us what happens if I switch this valve setting. And, in the network we have here, we basically also have the different partial correlations between the metabolites. So, we can actually say from our data, given that I switch the valve on a given enzyme, what would be the outcome on the rest of the network. And, that is very important information because if you would go into a theoretic pathway map, you would never know through which channels actually the metabolites flow and what would happen if I turn a certain valve because metabolism is a very reactive system. Only by experimentation you really know how the body would, what would happen if I shut down one enzyme by 50% or so. From the GWAS, that's actually what nature did for us. So, imagine the people with the, here. So, the people here, the red ones, they are all people where nature switched this thing on a low throughput. And then, you would have all the other settings in parallel, and if you go for thousands of people, you would actually end up with a full understanding that you could then use to rationally metabolize, to rationally engineer your metabolism. So, this is something my group is now working on. So, we got funding from Qatar National Research Fund on that. And, just very quickly, so the idea now would be to say we want to put up the, so we want to change basically pathological metabolic state. That can be diabetes. That can be cancer. We know where the valves are, and we know how we can attack them. And, many of these are already drug targets, so we know, For many of these valves, they actually have already actionable drugs to go for, and the others that we don't have them yet for could actually be even very interesting drug targets for use in this approach. And then, what we presently do is we try to develop the in silico models, systems biology models, to actually mine the network, and then to work with them, have in vitro models. So, at the moment, it's cell culture from cancer cells where we can test the drugs and the metabolic assays to actually look at the outcome. So, I won't get deeper into this, but I think there's a huge potential here to look for new drug targets and use GWAS results in order to modify metabolism in complex disorders, and, of course, also in cancer. So, imagine, for instance, a person who has something like a PKU, passenger mutation in the cancer cell, that would single out this one cell, and then, if you would shuffle toxic metabolites into a broken pathway in cancer cells, these would selectively suffer and be weakened and ideally die. All right, next thing, just very briefly, so I never personally worked on it, but I like the paper very much by Robert Plenger and coworkers. So, they thought about validating therapeutic targets through genetics. So, here you basically say genetic, if you're more than one genetic variant, you could actually simulate, and it's the same idea of using nature as an experimental laboratory. You would actually have different settings of your metabolites, and you could look for drug dose-response curves. So, he has put up in that review paper a number of criteria, and I just want to highlight a few. That apply to metabolomics. So, one would be the biological function of the causal gene and the causal variant are known. So, that's very clearly the case in our— in the metabolomics because we have this match between the enzyme and the metabolites. Then, genetic trait is related to clinical indicated target for treatment, so you know what enzyme you're looking for. And then, the causal variant is associated with an intermediate phenotype. So, that's really the key thing. The metabolomics is the intermediate phenotype that can be used as a biomarker. So, you can actually use, if you now work on a drug for diabetes, and you know you have an intermediate metabolic profile as a phenotype, you can optimize your drug in order to optimize the intermediate phenotype and not the endpoint, which would be much, much cheaper. Of course, in the end, you have to test it on the endpoint again, but in the intermediate process, you can be much more efficient. So, here's, for instance, an example how that would look like with two genetic variants. Okay, and I think then I come to the final part, I think, which is also most interesting now for the Precision Medicine Initiative in the U.S., and we have similar things in Europe and in Qatar, as well. So, a few years back, we analyzed data from the KORA study, where we had data from more than 800 people who were taken at 2 time points 7 years apart. And, what you see here is basically what we call the metabolotype conservation index, which is basically the correlation between the metabolic profile of a person and all others 7 years ago. And, it's the rank of that person's own metabolotype 7 years ago in the list. And, what you find is that about, for about 40% of the cases, if I take if I take the correlation between the 7-year distance metabolic profiles, I would actually be able, in a one-to-one way, identify the person who the blood came from 7 years ago. So, that means that there's actually a huge stability in metabolic profiles, and only 5% would really change drastically over 7 years. And, these 7%, they could not be spotted on PCAs. So, these are not outliers at some point, so they would be, right in the middle, so I wouldn't be able to spot them in each time point alone. So, what this means that these people really underwent major changes. So, in KORA, we cannot test what it is because there's— it could be everything. It could be someone getting healthy, getting sick, getting pregnant, getting old, everything. So, this means there is basically a way of identifying major changes in an individual by looking how— looking at how their metabotypes change. And, I think that's an approach that is taken at different places. I think also in the biological passport for antidoping, they take a similar approach. The idea is to say if this here would be the distribution of some metabolic marker in the population, there would be a normal distribution, but in order to make that a biomarker, the effect size of the association you're looking for should be must be rather large. It should jump at least maybe one standard deviation or so. However, with what we said before in the study before is we have shown that the personal variation is much, much smaller. So, the personal normal distribution is not in this width, so I could go and take the initial visits for maybe after a week, a month, 3 months, blood from the same individual, and establish his or her own personal normal distribution. So, I would basically expect that that person should stay in this range and not in the big range. And then, I can look at that person at continuous visits in the future, and I know, for instance, Metabolon has started to look at this with patients who are at risk of cancer, for instance, people at high risk of cancer. They would just monitor them and just say,. We just watch out for anything that's abnormal, and abnormal would then be defined on your personal profile. And then, if something jumps out, you can alert, get an alert, look at all the other markers, and you may find out that you have a certain combination that maybe not only one is out, but a certain profile, and one may be up, one may be down. You have the information on that. So, the thing is, the nice thing is we now have all these studies that we've done. We can basically pinpoint what these metaprototypes associate with. So, there's a lot of metabolome-wide association studies telling you the metabolic footprint of diabetes or oncometabolites or GFR. So, I could say this patient has a perturbation in his metabolic profile, and his metabolites look like, for instance, a kidney problem. So, I would refer him to a nephrologist and say, "Tell your nephrologist to look more deeper into your kidney." And, the same thing is if there would be oncometabolites, I would rather put him to an oncologist to check whether he might have somewhere a hidden cancer lump. And then, the idea is to refer that patient further to a specialist who would then take charge of it. So, we're not trying to be predictive here in a way that we don't need the doctor. It's to direct him to a doctor in a fairly non-targeted way. And, I think there are people doing this already. So, there is HLI. So, I think Craig Venture is certainly going to do this. So, here I have just some press releases. And, I think there's some interesting keywords here, like the access to the Twins UK registry together with Tim Spector, who was one of the leading researchers on our ATLAS paper. And, they will actually sequence 6,000 longitudinal samples from the Twins UK. With Metabolon, that's the one we've been working with, right? So, he's actually picking up now on the GWAS and trying to translate this into a real application to human health. And then, the same thing— not the same thing, but a similar approach is, of course, precision medicine. So, here again, the aim is to recruit really large numbers, and what I personally think is we're really standing at a new era of public versus private now to understand really what the genome really means. And, I think this is really a very positive thing because those of you who know the history of the human genome, in the end, we got the human genome probably 10 years faster because of the competition of public versus private. So, I'm very hopeful this will be a very positive and constructive competition, and I think the key here really is the deep phenotyping of large cohorts in order to get beyond what we've done so far. So, I think, I hope I have shown you a little bit how that works. And, just to finish things up, I have some resources here that you may look up later. So, one is I have a blog, metabolomics.com, where I try to follow up on different interesting parts of human metabolic individuality. There is a web server called GWAS.eu. So, on that server, we basically put all the GWAS association data from the different GWASes, so you can go there with different search tools and say, is the gene I'm looking for interesting. What you should remember here is in these GWAS papers, we always report the genome-wide significant hits, the strong hits, the ones where there's absolutely no doubt on, because we want to publish these and we don't want to speculate. However, if you're interested in a research project and you want to generate hypotheses, then you may be interested even going below the one-for-only level, because there will be a lot of true positives in the net, and if you have other information about a certain locus, you may be able to pick also true positives out of 10^-6, 10^-7 associations. And, in any case, it would be hypothesis generation, so you would do extra experiments to basically confirm that. So, it's really to guide experiments here. Another server we had is the snipper.org. So, this server basically also contains the GWAS data from the metabolomics, but in addition, it also has a connection with the GWAS catalog, with ENCODE Roadmap, with GTEx, and it basically integrates all information that we have on a deep molecular phenotyping level onto a variant which you would use to query the thing. And extend information through linkage, through correlation between SNPs, to entire blocks. So, the tool I really like at this server is the block annotation. So, I can go there. Well, yeah, I mean, the practical uses is you go to block annotation, you take from a new GWAS that you may have run a genetic variant, you paste it there, you run it, and it will tell you everything that may be of interest— also below bonfoni level, of course— that may be of interest in terms of metabolomics, expression level, regulation, and GWAS on these variants. So, it's really a very nice tool to annotate intermediate phenotypes to GWAS. And then, so everything I told you today, maybe it was a bit quick and everything, so recently we were invited to do a review for Human Molecular Genetics. So, here's the PubMed ID. So, in this one, You can read again what I presented today. And, I think, finally, I think this is something Rima Daouk from Duke is really pushing very hard, is in terms of precision medicine. She's animating a precision medicine and pharmacometabolomics task group, and within that task group, we started working on common protocols and things. We basically have a white paper on how metabolomics can actually enable precision medicine. All right, and with this, I come to the end. So, I should acknowledge my funding. So, most of it comes from Qatar Foundation under the Biomedical Research Program. And, of course, all that work hasn't been done by me alone, far from that. So, much of it has been done while I was in Germany. The group I headed there is now headed by Gabi Kastenmüller. So, these are the people at the Helmholtz Center. So, they're still very strongly working on on the Chorus study. The people around the GWAS, so here are all the names, the key people making this run. There's, of course, Tim Spector, who really brought in the funding and the samples for the big study. At that time, he worked very closely with Nicole Sorensen and So-Yeon Shin. She's the first author here. And then, my colleagues from Munich, Gaby and Christian, and the Systems Quality, especially Jan Kunzik. Finally, a special thanks to all my research team in Qatar, where we are now working on the things. And with this, I think I thank you. So this is our lecture hall in Qatar. And should you have any questions, here is my email. And I think with this, I can hand over and hopefully take some questions.

Thank you very much. Great. I want to thank Dr. Zurre for a really great and interesting presentation, and we do have a few questions, so I am going to start to read through those. The first question is, how to distinguish between noise and signal for genome-metabolome correlations?

Which tools and approaches help discriminate false positives versus true positive? All right. I mean, that's a really great question. We always, for the moment, we took the safe road. So, the safe road, you just go for one for only significance and replication. So, you basically only consider true positives, associations that are significant at a level where you correct very conservatively for all tests all SNPs, all metabolites, and in addition, which statistically you don't even have to do, but in addition, you replicate it in an independent study. But then, of course, you're losing a lot of true positives. I think that's more your question. So, below that level, I think you have to take some kind of a Bayesian argument approach, which I never formalized, but maybe other people may have done that already. To say, for instance, if a genetic variant is close to an enzyme and it's matching the metabotype, then actually I don't have to correct for 500 metabolites and 500,000 SNPs, but only for the number of SNPs that are around that enzyme and the metabolite that actually makes sense. So, that would be, I think, a less conservative, but still pretty safe way of getting a true positive. The other thing is, of course, always replication, so if a signal appears twice or even more often, even if it's not that strong, but if it reappears, especially if there are different techniques, different cohorts, association is always a bit dependent on the noise in your measurement. And metabolomics, high-throughput metabolomics measurements are not 100% precise. There are a lot of potential artifacts that can come from sample handling, from machine settings, from even from population structure, whether people were fasting or not, how the blood was treated, and so on. So, if a signal appears under very different conditions and it still appears, then I would say this is really a robust signal. So, this kind of signals I trust. What I don't trust, if it's a single strong signal that does not reappear in another study that actually has much higher power and much better measurement. But, in the end, it's pretty subjective once you go below the genome as a presentation level. Okay, thank you. Are we ready to move on to the next question? Yeah, I could go on for this, but I'd rather take another one.

Oh, okay, I don't want to interrupt you though. If you have more to say, please continue. No, no, no, that's all right, that's all right. Okay, wonderful. We'll move to the next question then.

How difficult would it be to conduct these association studies "Non-model organisms and smaller sample sizes in terms of approach and confidence on the results?" Well, I think that has already been done, and it can be even easier done because the big problem in humans is we have a lot of variation on top of what our genes do. So, most of the time, it's only the minor part of variation that is explained by the genes, and the larger part comes from what we eat, what we drink, how we are, and so on. And so on. If you go to model organisms, you can control all these environments, so the noise will become much smaller, so you don't have to do the same study in thousands of mice. And, I know, for instance, Marc Dumas from the Imperial College, he's been doing a lot of great studies in mice, and if you look up his publication, you will find a lot of examples in that direction. So, I think it's absolutely possible. I also have even seen GWASes on, hang on, what was it? It was on some plants. Wheat? Maybe wheat? No, rice. Rice. So, you can also take different varieties and identify variants there. It's always a question of power in the end, and if the effect sizes are big, especially if it's really a loss gene, then it's not really a GWAS anymore. Then it's basically looking for monogenic disorders. Then, you could even go and just, and maybe even in, well, I don't know, maybe 20, 30, 50 different individuals, you could find some Okay?

Great. Moving on to the next question. Okay. Would you be able to comment on the predominance of enzyme-based signals in the GWAS studies? Why enzymes instead of other biological effects? Do you think this is mediated through higher or lower levels of enzymes, or through production of enzyme isoforms with different activity levels?

Hmm. Well, first of all, I don't want to be dogmatic. Probably, it's all of them. What I personally think is, well, first of all, we've seen there's not so much overlap with eQTL, meaning that where you have a metabolic association, you also find an expression level different than from other things. But, what I think is that the expression of an enzyme will be regulated in the tissue, will be regulated by expression. That will not, would not explain metabolic association. So, I think most of the variants in metabolomics are probably due to changes probably in the protein themselves, binding activities to certain different classes of molecules, or the speed of the enzymatic reaction, and things so on. So, I think most of these on the enzymes are really should probably be structural, although I've seen other examples.

Okay, the next question. In your GWAS, if you were to look at genotype-phenotype relations in blood versus urine in the same person, how similar or different are the relations?

Can you use the difference to see the metabolic processes? Yeah. Well, we have recently started to do that, so there is an example with a transporter. I don't remember the name, but Johannes Ruffler and coworkers have worked on that. So, there we have basically found a genetic variant working with myoinositol, and it's the same genetic variant in urine. It goes the one way, and in blood, it goes the other way. And, that is very strongly pointing for a variant that would actually modify the throughput of the transporter. So, I think that's a very clear example of that. Then, you may have other associations where in urine and blood, it may go the same way, and then it would probably rather be a signal of basically the urine mirroring what is in the blood. So, if you have more in the blood, more ends up in the urine. So, the directionality of these associations would actually be also very informative. And yeah, I think there are GWASes in the works where they are basically working on urine and plasma in parallel.

Yeah.

Okay. I have a question.

So I know you've done some thinking about this, and I hope I'm not putting you on the spot, but I think this is a great forum to maybe talk about this a little bit, and you had mentioned sort of precision medicine, precision medicine initiatives, and the paper that came recently from the Metabolomics Society. I was just curious, you know, what considerations need to be taken into account at the point of sample collection for metabolomics to be effectively included in these efforts? Because I think a lot of times, you know, we hear precision medicine and we think genomics, but I don't know, I really think there's an opportunity here to do other types of omics, and I was curious what your thoughts on that is.

Yeah. No, very clearly, when we talk about precision medicine, I think genomics is only the first step of it. And then, of course, it becomes tricky, and I think in the discussions we had, especially to write this paper, you really get into this twist like, oh, I would have everything ideal, and Oh no, I want to do a million people. And I think that's really a question in the end of resources and to decide where do you get the best shot for the bug. I really think, well, we can learn a lot from other studies like the Framingham, the CORAS. My personal view is you should really go and make sure that you get enough alcohols, at least from the blood, and the urine is easy to collect, saliva is easy to collect. You don't want to miss the stool samples, but again, that could be already pretty, pretty complex. So, I would maybe— but that's my personal opinion. I would maybe not put so much money into that if I had the choice alternatively to put it into larger sample sizes. Oh, sorry. I don't know why the— Things are going off. Yeah. Right, so I think there need to be discussions a bit of how much resources we have. I understand at the moment, I think the different cohorts collect a lot of samples. For instance, at Cornell, I understood they have 150,000 they are collecting. And I think they still have to learn a lot what they could really do with them, and not just collect them for genetics. But, I think, yeah, let's go with what we wrote in the papers. I don't want to deviate from that. But, I think it's something we really need to discuss with the funders, how much money they have to do it. Urine, saliva, plasma, maybe 5, 10 aliquots per thing I would definitely go for.

Great. Thank you for your thoughts on that. I thought it might be important sort of to hear a little more about that since we have there's actually quite a few people on the call, and I think it's a topic that I keep hearing coming up again and again.

So, I do want to add— Sorry, that's just because I don't want to make a firm statement there, because I'm more the bioinformatician who likes working with the data. I think there should really be a consensus of the community, not just what I say, what I think. I would like to have, of course, as much data as possible.

Right, right. And I think there are a lot of discussions ongoing about this topic within the community, and I'd like to see that continue. So I fully agree with you. I am going to give people just a minute. Are there any further questions for Dr. Zurey? We have him for a couple more minutes on the call. Now would be a great time if you've been sitting on a question and not sure whether to ask it.

Well, otherwise there's my email, so you can always send it to me.

Yep. Okay, I'm going to just give people just a little more time in case they're typing, and then we'll move on. Okay, I think those are the end of the questions. I really do appreciate your time today and want to thank everyone for their attention and active participation in today's NIH Metabolomics Interest Group webinar. Webinar. I also really want to take this opportunity to thank our presenter, Dr. Zurek, for a really excellent presentation, and clearly there were really a lot of people on this webinar. You probably can't see it, but I can, and I think there's a lot of interest in this area, so we really do appreciate your time today to share your research with us.

And then we would welcome— I'm very happy, and thank you very much. It was really very, very much a pleasure. It's always a bit weird to be sitting in your kitchen and just talking to your fridge, But I hope people enjoyed it. So if you have any feedback, please email me as well, because I haven't seen any one of you.

Yeah, I believe it or not, you're not talking to just your fridge. There's quite a few people on this webinar, so it's really been, it's really been really excellent. I'm really excited about it. So again, and also we welcome anybody for feedback for the future webinars or any suggestions for speakers or anything. So with that, again, thank you all for attending. Thank you, Dr. Zurey, and we will sign off.

Everybody have a fantastic day. All right, bye-bye.