JAEYUN SUNG: Good afternoon, everybody. My name is Jae Sung and I'm an Assistant Professor in the Department of Surgery in Division of Rheumatology at Mayo clinic, as well as a member of the Microbiome program in the Center for Individualized Medicine, or CIM. I would like to welcome you all to the CIM Grand Rounds.

The CIM Grand Rounds lecture series is designed to highlight the latest in scientific discovery and innovation and demonstrate how individualized medicine is being translated into the practice to meet current and future patient needs. Today, we're highlighting the theme of beyond DNA, which looks to advance patient care by using molecular information from the microbiome, immunome, epigenome, exposome, and more. I'm enormously excited to introduce today's speaker, Professor Michael Snyder.

Dr. Snyder is a the Stanford B Ascherman Professor, Chair of the Department of Genetics, and Director of Genomics and Personalized Medicine at Stanford University School of Medicine. Dr. Snyder received his PhD at Caltech and carried out postdoctoral training at Stanford University. He was a faculty member in the Department of Molecular, Cellular, and Developmental Biology at Yale University for nearly two decades, rising to Chair of the Department before moving his lab to Stanford University in 2009.

He is a world-renowned leader in the field of functional genomics and proteomics, a pioneer in self-tracking personal health data through multi-omics profiling and wearable sensors, and one of the major principal investigators of ENCODE, which is an NIH-sponsored multi-institutional research project using advanced tools to identify all functional elements in the human genome.

At the Stanford Healthcare Innovation Lab, Dr. Snyder runs a team of over 100 scientists doing some really inspiring work that is essentially building the future of healthcare into a system focused on disease prevention. Dr. Snyder was elected a member of the American Academy of Sciences in 2015. And on top of all his scientific accomplishments, he is-- he co-founded over a dozen biotech firms in genetics, genomics, and precision medicine based on personalized big data, all with the main aim of keeping people healthy.

Today, Dr. Snyder will be speaking on new advances in health monitoring, describing how his lab's work during the past few years is radically transforming healthcare with deep multi-omics molecular measurements and wearable technologies. More specifically, he will summarize his recent findings using those technologies, including the concept of ageotypes, how wearables can detect COVID-19 and other infectious diseases, and remote sampling for health monitoring.

For the audience, we'll have a Q&A session at the end of Dr. Snyder's talk. So please feel free to use the Q&A in zoom to submit your questions. And for those who have not yet claimed continuing education credit, a message will be sent out in the chat shortly. Michael, thank you so much for accepting our invitation. The floor is now yours.

MICHAEL SNYDER: OK. Well, thanks, Jaeyun. It's really great to be here and present some of our latest work. In fact, one of the topics I'll present just came out today in Nature Biomedical Engineering.

So as Jaeyun said, our stick is to try and use big data to transform healthcare And let's get started here on the sharing. All right.

So I would argue our healthcare system, I'm sure many of you would agree, leaves much to be desired. And there are many ways in which we can probably improve healthcare but these are some of the elements that particularly bother me. We are very focused on treating people ill. As such, medicine today is very reactive.

We measure very few things, especially when people are healthy. Also, when you're healthy, you don't get measured very often. And one particular problem is nearly every decision about your health is based on population-based measurements. Instead, I think we would all agree it's better to keep people healthy, be very proactive.

We can certainly measure many, many more things than we currently do. The frequency, I would argue, should depend on what you're at risk for and what your personal trajectories look like. And lastly, we should be doing more individual-based medicine than we currently are doing. You know, we do do some, but one could argue we can do this much, much better. And the net result is we should be very focused on precision health rather than precision medicine.

So to illustrate this point, I'd like to show the slide here, which is your oral temperature when you put a thermometer in your mouth. You've been told since you were little, it's 98.6. Yet if you actually look at the studies that are out there, this is a very typical one, you see lots of credit here.

But basically, from a study of 3,000 people, the median temperature is 97.5. It's not 98.6. And mine personally is 97.3. It's actually been dropping over the last eight years.

I think more importantly, though, is there's a spread. You can see this is a 25th quartile, 94.6. This is a 75th quartile, 99.1. And what that means in today's world is that if your normal, healthy baseline is here, 94.6, you go to a physician today and they measure you at 98.6, they'll tell you you're normal. Everything's healthy.

But if you're up four degrees over your baseline, I guarantee you're not healthy. And so that's a big theme of our work. Know what your healthy baseline is so you can detect a shift and catch disease at its earliest onset.

So we're in a world now where I think we can start dissecting things that impact your health. So your health is influenced, obviously, by your DNA and your genome. A lot of emphasis is placed there. But we know intuitively, all these other things impact your health, exercise, food you eat, stress, various exposures, all impact your health.

And we're in a world where we can now quantify this. So you can quantify exercise, your activity monitors, food by logging, that one's still clunky, stress indirectly, and other things. We're actually ourselves working on a lot of technology, especially environmental exposures.

And so we can quantify many of these parameters. But equally importantly, we can quantify the impact of these parameters by doing deep molecular profiles on people. And so a number of years ago, we set up something we called the Personal Omics Profiling Project, where, in addition to sequencing people's genome, we actually made many other measurements, mostly out of blood.

So we'll follow their DNA methylome transcriptome proteome out of peripheral blood monocyte cells, immune cells, cytokines and proteins out of plasma, metabolites, lipids also out of plasma. We do follow the microbiome. We do gut, nasal, and more recently, we just completed an analysis of tongue and skin. That turns out to be quite interesting. And these clearly impact your health.

This is on top of deep clinical tests, questionnaires, stress echocardiograms. These are more advanced tests, and a lot of glucose control tests, because it turns out I'm type II diabetes. We'll talk about that in a minute. And so we do oral glucose tolerance tests and some resistance.

And then about eight years ago, maybe a little longer before we started getting involved in wearables. It was biosensors, and so added that on as well. So one aspect of all this is to collect deep, deep data on people, and the other aspect is to do it longitudinally. So we'll sample people every three months while they're healthy, and then if an adverse event comes along, say, a viral infection, we'll take typically five to seven additional samplings.

And so we've been running this on a smallish cohort, about 109 people. Some are now dropping out and we're adding others on. And we've been running the study for-- well, I'm one of the participants. 13 years on me. It's almost 10 on the rest.

So you might say, why are we doing all this? And the answer is we're trying to understand what it means to be healthy. What does a healthy profile look like? How does it change over time? How does it differ between different people?

What happens at the earliest times of illnesses like a viral infection? And lastly, can these advanced technologies like genome sequencing, these omics technologies, can they be better used to manage people's health? And with regards to the last question, omics technologies and health, we actually discovered that from the smallest cohort, nearly half the people learned something important about their health.

About 47 people signed up. Two issues were found, and it spanned a wide range of areas. Heme/oncology cardiovascular disease, metabolic, so on and so forth. So in each of these areas, people learned something pretty important about their health, all in advance of symptoms. So that was what's important about this.

By doing these deep data collections, we could see when things were off in advance of symptoms. And it was no one technology that did it. Sometimes-- we'll talk about this more in a minute. Sometimes it was genome sequencing, sometimes it was imaging.

And some of these were a big deal. We didn't catch someone's early lymphoma, two people with precancers, some with smoldering myeloma. Two people with serious heart issues, one from genome sequencing, one from wearables, and so on and so forth. So again, all these were found pre symptomatically.

Now the way we like to think about it is that if-- the way we do medicine these days is taking eight pieces of a Jigsaw puzzle to try and figure out what the picture on that thousand-piece puzzle is, we're actually trying to take more like 600 pieces where we'll get a much, much clearer picture of people's health. That's really the idea. And so again, quite a bit was found out.

So just to give you a few examples, genome sequencing was powerful. 12 people. Actually, 13, if you count my polygenic risk score, and actually 14 for one of the pharmacogenetic players.

But 12 people from the first 70, it's a little bit higher now, learned something pretty important, just from their genome sequence. And these are what we call Mendelian mutations. These are single-gene mutations that, by themselves with a high incidence, can lead to disease. And of course, the most famous one are the BRCA mutations.

So we did have an individual with a BRCA1 mutation. That person was an orphan and learned that then they were at risk for breast cancer. And so again, we found a number of these.

It turns out three of them actually wound up having impact on disease, meaning this individual here, after we sequenced his genome, we saw that he had a mutation in a cardiomyopathy gene. And it turns out his father died of a heart attack at around the same time. Aunt had had one in her 50s.

And then he did a stress echo. And as a consequence of this genome sequencing, sure enough, he's got a heart defect. He's on drugs now. So the point out of all this is that we could actually uncover this from genome sequencing in advance of symptoms.

There's another individual-- oops-- who had a mutation. And this gene puts them at high risk for certain kinds of cancers. They did a whole-body MRI follow-up, had early thyroid cancer. That was removed. They were able to keep their thyroid, didn't need thyroid replacement therapy.

And the net result is-- again, that was all caught early. And both these folks were quite young, actually. So it was a big deal. I forgot to say that the cohort as a whole was 53.4 when they entered. So that's probably why we found so many. It is an older cohort, but it's a wide spread. So we did have young folks like one of these here. 2x means there were two individuals. Even some had mutations in APC pathway.

There were nine individuals coming into the study thought to be type II diabetic. And one of them actually turns out that mutation in this gene, they're MODY diabetic, which is a different form many of you I'm sure know. And the way you treat a MODY is often different from that of a type II diabetic. So they've actually got on better medication now. They've been on a suboptimal one for years.

So the net result, again, is that even though it's a minority of folks who learn something valuable from their genome sequence, if you're one of these, it's a pretty big deal, like the BRCA person and so on and so forth, the heart person. So we think genome sequencing can be quite powerful.

Many of you may know that most forms of disease, when you think about genetics, they're really not due to these single-gene mutations like BRCA and stuff. They're actually due to complex disease. And most complex disease, you may or may not know, it's thought to be due to these common variants, these common changes in your DNA that are thought to be a very small effect. And people who are unfortunate get a lot of the bad variants for these things, and type II diabetes, rheumatoid arthritis, all these things. They're the most common forms of the complex disease.

And so what people are doing-- so you can interpret these single-gene mutations the way I just described. You see a mutation. If it's a known pathogenic variant, you can then coach folks.

For complex disease, what people are doing-- oh, I forgot to say. The analogy is that this is like clobbering a gene with a sledge hammer, and therefore leading to disease. And this is more like death by 1,000 cuts. Lots of little changes. A small effect basically caused the problem.

So the way people are dealing with complex disease these days is using something called a polygenic risk score. So we're going to do a touch of genetics here. And in polygenic risk scores, what people do is they sum over these variants.

They have an increased odds or decreased odds of causing disease. And these days, people will sum over thousands, if not millions, of variants. Here's a paper from a few years ago looking at these five, again, complex diseases, some over millions of variants.

And this sort of works. It works for the top couple percent. So if you're in the top 2% or 3%, you can actually get pretty good predictive value. But if you're below that, forget it. It doesn't work that well.

And we think that there's two reasons for this. One is that the way they do this is by linear summation. And genetics isn't linear. For any of you who are geneticists, you probably know that one plus one is rarely two.

There's interactions like one plus two mutations in the same pathway. One plus one there would be one, because you won't get additive scores. And if you're a different pathway, sometimes one plus one can be 15. That can be very additive, but more than additive effects.

And so that's all missed. And the other thing that's missed are all the rare variants that contribute to this. So this only assumes these common variants.

So we came up with some new methods for analyzing genomes that have been quite powerful. And I'll tell you about one of them in a minute. They both use machine learning, which I'm sure you appreciate, is really transforming a lot of different fields.

When we first applied this to looking at a disease. In this case, I'm going to give you the rare variant example. I'll show you-- I'll just summarize a common one in a minute. But we looked at a disease called abdominal aortic aneurysm. You may know, but this is actually a big killer of people over 65. It's actually the 10th leading cause of death.

And unfortunately, the way it's discovered is that-- when people's aorta bursts. Then they die 90% of the time. So that's a terrible way to know you're at risk for the disease, obviously. It's known to be highly heritable. 70% heritability, and there are some markers associated with this.

So we wanted to see if we could try some new machine learning methods to get at the genetic basis of the abdominal aortic aneurysm. And what we did was something that would just totally-- and it did at the time when we published. It just totally mind boggles GWAS folks. So if you're a GWAS person, you're probably going to say, hey, there's no way this is going to work, and it works great.

But the idea is the following. We sequence-- we only have four-- we got 401 great genomic sequences. 268 cases, 133 controls. We did do whole-genome sequencing, but we're really just calling pathogenic variants inside the gene. So we actually did a decent depth of sequencing as well, 50x.

And then we did this machine learning where you basically say, give me all the pathogenic mutations. Give me the genes that are enriched with pathogenic mutations and cases relative to controls. We also-- you can mix in other features while we mix in the limited amount of EHR data.

And wouldn't you know it? At the end of the day, we came up with 60 genes. There had only been three genes known before, and even those were kind of sketchy. We found 60 genes underlying this case that were hyper rich for pathogenic mutations, relative to control.

You do, for the aficionados, 10-fold cross validation, and can show all this. And it's like a 10 to the minus 13 p-value is what comes out of this. And then we went on-- I won't have time to show you this, but we can show the pathways that the 60 genes are involved in. They're exactly what you would expect. Blood circulation, blood pressure, cardiomyopathy, aneurysm.

We also looked at mouse models for AAA. And sure enough, the paralogs of this, they're all involved in cardio and these aneurysm kinds of functions. And lastly, we could show the genes are missexpressed in both A patients and in the mouse models. I won't show you all that data, but there's a lot of validation around this to show that these 60 genes were, in fact, the right genes.

And then we could even go on and build a predictive model that actually we think could be used clinically. So we're trying to test this out. But basically, the genome alone actually has reasonable predictive value, 0.7. Electronic health records can be valuable, especially later in life, because they'll show high cholesterol and can be predictive. But the two together even better. I know that looks modest, but actually, that's a pretty big jump in sensitivity.

So at the end of the day, we now have a predictive model. And I as we add more and more data, we'll get even better. And so that's what we're trying to do. And the models are as good as certainly the coronary artery disease models that are out there now from Framington, and the more recent versions of that. So we think this is going to be quite powerful.

We've now gone on to actually do machine learning in other ways for two other diseases, some ALS, which many of you know, it's also known to be highly heritable, 61%, and severe COVID. So young people get COVID can wind up in the hospital and that's thought to be due to genetics.

Here, we did a little bit different approach. We published this. So feel free to-- where I can share the slides. You can look this stuff up. But basically, there were seven genes. In the biggest study out there ever done, they found seven genes.

And with their machine learning, we can actually search the open chromatin regions and look for enrichments in GWAS signals in these open chromatin regions relative to control. So in that case, we focused on motor neurons.

At the end of the day, we identified 690 genes. So we-- you know, basically 100 times more genes. And we went on and validated several of these, and in fact, we validated all of them by many methods, misexpression, all the same stuff I showed you before. .

Very, very powerful. And-- excuse me, we knocked out one of them, tank one, which also turns out has a Mendelian form. Those were discovered as we were doing this. And sure enough, that has all the phenotypes of an ALS.

So the net result out of all this is that these methods are very, very powerful. Same with severe COVID. We meant this case, we used open chromatin regions from single-cell data and can look for enrichments across all 19 cell types.

And we got over 1,000 genes involved. We can explain it's a whopping 77% of the heritability. That's unheard of. It's amazing. We can really pin down the genetic basis of severe COVID. That's because it's this giant GWAS study.

So again, the net result is that-- and that's true for everything we're testing now. This is just a very, very big enrichment in finding genes. Now, we don't yet have predictive models for these. We're working on that. I'm pretty confident we can get there.

I'm happy to talk about that later. But the net result is these machine learning methods, they work for two reasons. One is we restricted the search space for open chromatin, Sorry, a little jargony for some of you. And the other is the fact that we're doing machine learning, which is nonlinear, and the two are quite powerful.

All right. So back to the cohort. Well, one of the things we did learn is that other technologies are valuable, too. Imaging was valuable. We caught some of the early lymphoma because they had an enlarged spleen. They had several blood markers that were up.

We caught some with this precancer called MGUS. Many of you may know it. It's when people have too much IGM and too many IGM-producing cells. Now this one did come out from, basically, the cohort analysis, meaning we had one individual, a pretty young person, had just way more IGM than everybody else.

It was 1,000 samples. Two samples were just way higher than everyone else's. So it was quite revealing. And in follow-up said, yep, got this MGUS, and so get monitored more frequently as a consequence.

We have found other discoveries as well. Just tell you one last example, which is type II diabetes. This is one that interests me a lot because I mentioned I am type II diabetic.

And so basically, coming into the study, I mentioned nine people were thought to be type II diabetic. They're the ones in blue here. Well, two were diabetic who didn't know it. That does happen. And a lot of people are prediabetic who didn't know. That's very, very common.

And then we're very interested in how do people become type II diabetic. Do they just gradually get there? Does something trigger it? How does that work? And in the first four years, nine other individuals became diabetic and a few others had diabetic range values weren't officially diagnosed, and others became prediabetic.

So with regards to how do people become diabetic, the answer is, of the nine who clinically were diagnosed, seven of them just gradually got there. So here's two examples here. Glucose is in orange, hemoglobin A1C is in gray, and you can see this person gradually just went up on both of these. And then actually, an intervention, I'll talk about that later, brought them back down somewhat.

Here's a person who got there from oral glucose tolerance test. Was in range, although nearly there. And then just got there gradually. But of the seven that did this, so seven gradually got there, two actually gained weight, and that's the classical way to become type II diabetic. The other five did not gain weight.

Now for two other people, it actually looks like something triggered it, and I'm one of the two. This is me and eight years of data. Like I said, I have 13 years.

Now in my case, I first became diabetic after a nasty viral infection. I was actually getting measured for an insulin resistance test and surprisingly, my fasting glucose was high, and I had elevated hemoglobin A1C. I got tested a few weeks later and sure enough, it was basically crossed the threshold. Officially classified as diabetic.

And in my case, then, I did a lifestyle intervention. I basically cut out all sugars. Had doubled my biking. I'd been biking some, but basically cut out all the excuses. And I started running and I gradually brought it down.

Now you may say, well, that's a little bump, but that's actually about 10 months to a year before I brought it down to normal range values. And so I was able to get it under control by lifestyle changes. Now it was running so boringly low that I stopped looking at it.

I did the perfect control experiments. I wasn't looking at the data. And someone else noticed here a few years later that my hemoglobin A1C was back up high again. I said, oh. I was surprised. I went back and looked.

I forgot to say. So this first one came after respiratory syncytial virus infection, which these days, is running rampant. Kind of interesting now, catching it. Nasty reputation. I got pretty sick when I had this. This was back in 2011. Or was it 2012? 2011, yeah.

And when it spiked up again. So we went back in a retrospective look and sure enough, it spiked up again. And two things happened. I stopped running, and I also got a second viral infection.

So it's not clear what triggered this. But nonetheless, it's pretty clear I have a glucose dysregulation. I forgot to say it was predicted from my genome, from my polygenic risk score. And so I'm at the extreme end. That's why we could predict it.

Now I did start running again. I brought it down. I never got it all the way down to baseline levels, but I did get it pretty far down. And then I was running as much as my schedule could handle, four days a week, four to five miles a run. And gradually, both my hemoglobin A1C and fasting glucose kept going up and up.

And so I switched from running to weightlifting. You can't tell, but I lift weights. And the idea is that weightlifting and muscle mass is better for glucose homeostasis.

And so I was hoping that would bring it under control. That failed. I did gain 10 pounds, I measured myself, of muscle mass. I do whole-body MRIs. But unfortunately, I just kept going up and up. And that was a little disappointing to me, to say the least.

So finally, I bit the bullet and did what all my MD friends were telling me. Mike, you have to take metformin. So I did that. And guess what? I'm a nonresponder. I just kept going up and up. I went all the way to 75. It's not shown here.

And so what was the solution for me? Well, the answer is more data. And it's pretty clear at the end of the day, first of all, obviously, I have a glucose dysregulation defect. I actually make insulin fine, and it turns out I'm insulin sensitive, so my cells respond.

So what's wrong with me? Well, I actually don't secrete insulin from the pancreas. I have a very slow glucose disposition. And we went on to deduce.

So it turns out I'm a good responder to repaglinide, which is a PRANDIN derivative that promotes release in the pancreas. And then that works. And then I switched to an SGLT2 inhibitor, because I have cardiovascular protected, the reason that's valuable for me. That also worked.

And more recently, though, they worked, but didn't get me low enough and I kept creeping up again. So more recently, I went on the GLP1s, and that's like two weeks ago. And that seemed to really work amazingly well.

So the net result out of all this is what we call precision diabetes. By better knowing what was wrong with me, we could prescribe the right drugs. And a lot of the stuff is very predictable from the data, if we would only run the tests for folks out there. So I honestly think we can use this kind of information. I know what we're showing is a research project, but we can easily adapt this to the clinic.

One of the most important things we learned from all of this is that people have personal profiles, and they're actually very stable, and they're also very, very personal, but individualized. They're different from one person to the next. So as an example, I'm showing you data from 12 people who had at least 10 or more healthy visits. I

Have six years of data. I'm the light blue one here. And no matter what ome, or test you use, people cluster according themselves. Like I said, I have six years of data.

All my spots-- I'm the light blue one clustered together. Here's the red person, light blue, purple, so on and so forth. People do cluster on top of each other. Even this transcriptome will do the same.

And so again, these profiles are-- they're personal. They're stable. And even if you undergo a perturbation like a viral infection, in this case, people ran to the VO2 max.

They actually shift their profile. Half your molecules will change if you run to your VO2 max. But at the end of the day, you will still look more like you than the person sitting next to you.

So I did the experiment twice on these two, one's the dark brown, the other's the yellow, and mine-- a year apart, mine jump right on top of each other. And then here's another person and so on. So the point is that the differences between individuals is greater than the difference between a perturbed state like running to the VO2 max, or even getting a viral infection.

It's not 100%, but it's mostly there that you do shift when you get these perturbations, but you still look more like you than anyone else. That's a really, really important concept, because if I want to tell the difference between a healthy you and a sick you, I'm wondering if you're sick, it's very hard to do that by taking sick measurements and comparing it to just a lot of other healthy people. But if you compare it to their baseline, it's a very easy thing to detect. So we're big proponents. Get your healthy baseline now and find that shift.

So the other thing we can do, I mentioned most molecules are pretty stable. Well, there are some that do shift over time. And so what we've discovered, we look for patterns of things that might shift. We found that actually, there are seasonal patterns in the data that turns out to be very, very interesting. But we also noticed these changes over time.

There's 600 molecules of microbes, microbiome microbes that will shift over time. And what we discovered is that people actually age differently. The different molecules that do shift differently in different people. So this is me, about four years of data. Here's another person with about three years of data or less. We think with as few as five measurements, we can tell how you're aging.

And what we discovered is everyone's aging differently. So this is me, my coagulation, metabolic, other pathways go up with time, in that very typical pattern, actually. But you look at this person, their top pathway that's changing over time is their cardiac hypertrophic signaling pathway, their cardio.

So this person's a cardio ager, and so it's due to do-- so we think their heart's off. Basically, we later learned they're stage II hypertensive, and that sort of fits with this idea of cardiovascular function's off. So what we discovered, then, is-- so we had 43 people with enough data. We said, how many patterns are there in the data?

And at the end of the day, there were four patterns. We call these ageotypes. Four aging patterns, kidney, liver, metabolic, and immune. So each column's a different person, each row is a different age or type. And the folks over here, these four are aging in all four categories.

This person's aging in three of the four are not much of a kidney ager, but aging the other ones. This person's kidney. This is kidney and metabolic. I'm in the middle age. I think this could be me, actually. Probably that one.

Not much of an immune ager, but age more than the other thing. So end of the day, we can see how people age. It turns out the information is clinically actionable. There are markers associated with these age types. I'll just show you two examples.

Here is hemoglobin A1C, this is diabetes marker. The group, as a whole, goes up. And for individuals, the dark ones, that's statistically significant. They go up as well. But interestingly, four other people actually went down.

And this is a simple one to interpret. They did an intervention. So two lost weight, two went on a diet, one started running. And so that undoubtedly explains this shift.

There's another one with creatinine. And here, what's interesting here is a marker of kidney function. The group as all goes up, as expected, but 10 individuals went down They're statistically significant, and eight of the 10 are on statins. So there might be something interesting there about statins.

But there still are many things to be desired about how we can do healthcare better, and I'm going to talk about now remote monitoring. If you think about what we do these days, people, they travel to the doctor's office, they go to-- the office looks a lot like one or four years ago. Mayo ones are nicer. I've been there.

But nonetheless, many of the technologies are the same. A few of them have shifted. A few more gadgets, but they're not-- they haven't changed that much.

Then they draw lots of blood from you. Sometimes it hurts. And then they run a few measurements I mentioned before to see what your profile looks like.

So we're believers that this whole thing can become a lot more done at home. And so we've been doing a ton with wearables, which I hope to get to in a minute. And also most recently, the paper that came out today talks about our work doing microsampling from tiny drops of blood, the kind of thing you learned about from Theranos.

And here, the idea is that you take a tiny drop of blood. We're using Mitrix and another thing called Tasso Devices. And we spent a lot of time-- we spent six years trying to find the right technology. And so in the end, we thought these were very good devices for collection.

You Fedex a sample to our lab, the lab processes it, and then we can do these deep profiles. Lipidomics, metabolomics, proteomics, and very targeted assays as well. So again, the idea is sampling at home and then the analysis kit gets reported back.

And so we first showed that most of the analytes we measured are stable. That was, again, part of the whole testing process. We put samples, we took blood from two people. One of them, put them at three different temperatures, along with storing them at minus 80 right away, and put them at different lengths of time.

It turns out there's a pretty massive experiment to look at stability across time and temperature. And then we also-- and we scored this. And at the end result was that-- this is duration, this is temperature, this is something called an interaction term, a combination of the two. And nearly all the analytes are stable to time and-- so 128 proteins. Most are stable. A lot of these are targeted assays.

Same is true for metabolites. Most are quite stable. And lipids, a lot are off.

Half of them are off. And we don't think that's a big deal, because you can actually just work quicker, and in the long run, you could actually measure exactly what's going on. So the net result is that most-- this assay generally works for most analytes.

And then we can also correlate-- remember, a microsample is whole blood. And we tend to compare this with plasma. So we did do that for the exact same person the same day, the microsample in the blood. And sure enough, there's a pretty good correlation in these analytes between whole blood and plasma. And again, there are different assays.

For lipids, it's almost a perfect correlation. For the metabolomics, it's more off. But most of them are spot on line, but some of them are a little bit more different. So anyway-- and we can correct for that, if we want.

So we went on to use this in two different ways. I'll tell you about two studies. Probably going a little fast. I hope it's not too fast. But it's kind of fun, these studies. I'll summarize the main points.

We decided to try this out by running what we call a shake study. Now you probably know that everybody responds differently to different nutrients. But exactly how? And it isn't so clear.

So we had 32 people. We wound up getting a full data set from 28. Sometimes, people miss a spot. What they did was they drank this Ensure shake, which is a complex mixture, and then we measured them at baseline, 30, 60 minutes, 120, and 240 after drinking the shake.

As I say, it's a very complex mixture. And then we can go in and see what the response is. And it turns out, you know, 221 molecules did shift after drinking a shake. And there are different profiles. Metabolites show up in the blood, early, lipids, a little bit later, acylcarnitines go down when you bring in all these other exogenous nucleus. So that's a third-- there are three categories of molecules here.

And if you look at some of the key molecules, this was some of the things involved in insulin response, you can see we can measure all these things from this tiny little-- they're 10 microliter microsamples. And one of the cool unexpected results was that leptin actually went down and some of these inflammatory things actually dropped with the Ensure shake. I'll come back to that in a minute. It turns out that-- you know, I'll come at that in a minute.

So one of the things we discovered is that different people have different responses to the shake. And so this is the carbohydrate response. Generally, the class of molecules behaves the same for any given individual. So the carbs, for example, all behave the same, amino acids, the same, so on.

So here's an individual. Look at their time courses. These vary. 30 minutes, 60 minutes, 220. This individual, their carbohydrates dropped after drinking the shake. Same with this one over here. Pretty dramatically, actually. There's a third one over here.

Then you have other individuals like this one. Their stuff skyrockets. This one, that one, so on and so forth. So, so different people are responding very, very differently to the shake. And we can go on and score all this. For simplicity, we just picked six categories.

Gray is the average score of the group. And then I'll just show you a few examples. This person over here, they're elevated for inflammatory markers, inflammatory proteins. They're decreased-- or sorry, they're increased for free fatty acids, but decreased for amino acids.

And then this person's quite the opposite, they're actually anti-inflammatory. A lot of inflammation markers drop in response to the shake. Free fatty acids go up. See, this one goes up, too, amino acids. So people are responding.

We teased out five categories of people and these are just some examples. Let's look at this guy. Their amino acids go way, way down after drinking the shake and inflammation up. So, again, people are responding very differently.

And we think this is a big deal, because if you-- 10% of people have inflammatory bowel syndrome, and if you can actually identify that and who they are, you could actually modulate what they should be eating and drinking. And so I think we can start pairing diets to their effects on inflammation, other markers based on this stuff.

The other thing we did was we wanted to see if we could do detailed profiles on people. So we picked a person. Turns out it's me. You can tell I didn't make this slide because it's my picture. But we did very, very detailed profiling on me for just over a week, a little over seven days.

We collected 98 samples. So roughly almost one an hour across all waking hours, including one in the middle of the night. So we're doing this dense sampling, and then dense profiling for proteomics, metabolomics, lipidomics, and targeted assays.

It's also wearing a smartwatch at the time and a continuous glucose monitor doing food logging. So you can really relate all these variables. And the net result is you can start teasing out the patterns.

So here's glucose monitoring, here's heart rate. I'm a type II diabetic, remember, so I will spike. Steps, food logs, so on and so forth.

These are the molecules that are showing a shift. Certainly, a majority of them do show some sort of shift.

And so here's a sanity check. This just shows that basically I drink a shake every morning. Two components in the shake, they go up, come down.

That makes a lot of sense. Take a baby aspirin. Used to take it every day. I missed a few days in this particular sampling period. So you can see my baby aspirin there. But now you can actually determine its kinetics, its metabolism kinetics.

And so one of the things we learned that's kind of fun is that even though I stopped drinking caffeine and coffee by noon, and even though I did that, I still had a negative correlation between caffeine levels and sleep. So the net result is now I drink less and I try and stop before noon, if possible. So I better get my next soon. And we'll see if that actually improves things. I should have done that.

We found lots of molecules-- we're behind, so I'll skip over this. But we find lots of molecules with a circadian pattern. Some of them, we think are associated with food, like this one, but others probably are not, or don't seem to be based on what folks ate. So there's some intrinsic-- lot of circadian molecules here.

We're very interested in causal relationships between activity, and biochemistry, and between biochemical markers as well. And so we invented a method to do time-lagged correlation. So some events may not happen right on top of each other. Step A happens, and then that causes step B. So you get these lagged correlations, if you will, between events and, say, biochemistry.

And so just to illustrate this point, you can look for a correlation between steps and heart rate. So steps, we know increases your heart rate. And so we could do that. We basically took steps, and heart rate, and did this lagged correlation association analysis.

And the net result is there's a one-minute lag between steps and increases of heart rate. And that pretty much makes sense. So you get a fairly rapid increase.

And then we could do that for all the biochemical markers, all the physical activity. And what we wound up finding is there's thousands of associations between heart rate, steps, and CGM, were at least over 1,000. I guess 1,200. Sorry, I exaggerated slightly here.

And these are some of the top ones. CGM, that's glucose levels, heart rate, steps. These are metabolites, lipids, cytokines that are correlated. But it's kind of interesting.

The cytokines-- the steps and heart rate show a lot of overlap. That's not surprising. But they also have differences, because there are certain things that trigger your heart rate that won't trigger-- that aren't associated with steps. So yeah.

And then we can see the pathways that change. This is just doing the pathways that correlate with heart rate. So caffeine is number one. Here's a nutrient signaling pathway, and so on and so forth. Some of these don't-- I don't fully understand them. Maybe some of you do. Cushing's syndrome correlates with increased heart rate, whatever that pathway is.