Summary

In this session, Michael Snyder, Professor in Genetics at Stanford University, reported the latest findings from their longitudinal trials of healthy people being profiled using various omics and wearables. The first part was focused on the necessity of collecting a healthy personal baseline, because the differences between people are much larger than significant variations from personal baseline. Second on so-called “ageotypes”, patterns of aging that were found in clinical markers collected over time, clustering aging in a way that’s more actionable. And the last part was about COVID in the context of wearable data and how they are able to predict the onset of infectious diseases and likely also much more just from this easily collectible data.

 

Presentation: Michael Snyder

I will talk about my favorite topic, which is using big data to improve people’s healthspans.

Medicine is broken in many different ways, we should focus on precision health rather than just precision medicine.

Here’s an example of a misconception that most of us have based on population based measurements – that oral temperature is supposed to be 98.6F. The reality is that the median is more like 97.5F, but the more important concept is that there exists a spread. In today’s world, if you have a normal baseline of 94.6F, and you get measured at 98.5F, you are told you are fine, even though you are probably not, because of the 4F difference. But for that you need to know your own baseline, not just population average or median. So that is a theme that you will hear multiple times in this talk – you should really know your baseline, so you see deviations from it.

Health is influenced by many things, which we can quantify now, and it improves over time. We can quantify some of them better and worse, some of them through deep molecular profiling, which is important when investigating interventions.

We set up something called personal omic profiling, where the idea is to collect big data on people’s health using advanced technology – we’ll sequence your genome, and take lots of other molecular measurements, like methylome, along with other basic measurements. The idea is to profile people when they are healthy and then see the changes when something happens.

During the first 3 years of the trial, we discovered 49 major health discoveries, 77 if you count hypertension, in a wide range of areas. Some of them were a pretty big deal, we caught lymphoma, pre-cancers, smoldering myeloma. It was no one technology that did this, sometimes it was genomics, sometimes blood markers, and sometimes wearables. Together they were extremely powerful, more than just one modality alone.

Genome sequencing was important, twelve people had one of the so-called Mendelian diseases, and they learned something very important about themselves, for example high risks of specific diseases, suboptimal medications, or even led to early detection of cancers.

Imaging turned out to be powerful in at least one case of early lymphoma – when the combination with elevated blood markers was very powerful.

We detected plaques in the arteries, which led to recommendation to increase statins.

We’re very interested in Type 2 diabetes as it concerns me as well. 9 individuals were thought to be T2 coming into the study, but it turned out that 2 more individuals were and didn’t know it, and a lot of folks were prediabetic – 9/10 pre diabetics don’t know about it, and 70% of diabetics will go on to become a full blown diabetics, so catching it early to be able to better manage it is useful.

 

We’re also interested in how you become T2 diabetic – do you gradually get there or does something trigger it? That’s very relevant to the current COVID pandemic. We followed that in our study, where 9 people became officially diabetic. It turns out of the 9, 2 people got there through weight gain, which is typical. But the other 7 didn’t gain weight. 5 of the 7 just gradually became diabetic, not clear why that would happen over time. They got there by different measurements, and we’re actually working on that because we think we can subtype different types of diabetes based on that. 2 other people spiked their way to diabetes after viral infections. That was my case as well. There is probably a genetic predisposition that gets triggered by the infection. I was managing my glucose well with running after the first infection, but then I got another one and stopped running and that’s when it returned.

Investigating my own data with further profiling, it turned out that I make insulin just fine, and my cells respond to insulin, but I am very slow in my glucose disposition, I don’t release insulin from the pancreas – and there are drugs for that, which I am a pretty good responder to that, while I wasn’t to Metformin for example. We call this precision diabetes, by knowing better what’s wrong with you, you can prescribe the right drug for the right situation. And this could be done on anyone in a fairly inexpensive way.

One of the biggest flaws of the system relating to aging is that people all have personal profiles that are all very different, regardless of the measurement used. We’re all different, which you can see on the dots in the charts, where each color is a different person. The dots from the same person tend to cluster together, and the clusters are quite stable, we think these profiles are pretty robust and hard wired.

Beyond that, if you get hit with a viral infection or some other perturbation (in this case people ran to their VO2 MAX), you will shift your profile, about half your molecules change, but you will still look more like yourself than anyone else. So people have these really robust profiles that don’t shift, which is important because the difference between healthy you and diseased you is usually less than the difference between people. So if you are trying to see when you first get ill, you won’t pick that up very easily if you take a measurement and compare it to 500 other healthy people, because the difference between people is very very large. But it’s very very easy to see sick yourself from your personal healthy baseline, if you have it. So we’re big believers that you should get your healthy profile. And that’s an important basis for what I’m going to tell you next.

By the way the data is entirely out there in public, so anybody can do whatever they want with the data, you’re welcome to take a look at this. There is some fun stuff you can do with this data, one question that appeared was how many seasons are there really in biological terms. People say there are 4 seasons, but how do we know, right? So we looked at the patterns in the data in Northern California.

Most molecules don’t change, but there are about 150 that do and show some seasonal patterns. Some of them are known, but most of them were not known to be seasonal before.

So we can then look at them and see how many basic patterns they fall into. And the answer is two major patterns. There are some molecules that don’t fit, but most of them do. Then of course you can say that it’s Northern California – it makes sense, one should be winter and one summer. You would be half right, in pattern 2 you see molecules that go up in winter, related to different problems. The other pattern (1) turns out to be late April/May. In hindsight it makes a lot of sense, because that’s when allergies are pretty rampant, but there are also other molecules related to different problems that are peaking at that time as well. We think some of those molecules correlate with activity people exhibit during those times as well.

There are some clinical and microbial markers that change with seasons, which we think is relevant, because you can take this into account when you are starting to see molecules change (for quantified self folks for example).

Here’s what is most relevant to aging – since we follow people over long periods of time, we can look at what’s changing over time. We did the classic comparison of old and young, where we got expected markers and some new ones because we are profiling so deeply, but the more interesting part is how they are aging. We have 43 people with enough data (5 measurements within 2 years gives a pretty good profile on how people are aging). It turns out that everybody is aging differently, which isn’t a surprise but hasn’t been really shown. Person 1 is a typical ager, coagulation and metabolic markers, they all go up over time. Person 2 however has the biggest downward change in cardiac hypertrophy signaling pathway – so they are more of a “cardio ager”. In hindsight we looked at them and they have stage 2 hypertension which sort of fits with this concept. Since we have 43 people, we can ask a similar question as with the seasons – how many patterns are there that these 43 people fall into?

The answer is 4 major patterns. Now I know that there are more than 4 major patterns, but those 4 we had statistical power for. We call these ageotypes – aging patterns if you will. Kidney, liver, metabolic, immune. Each column is a different person and each row is a different ageotype, and you can see the people at the right end are aging in all 4 different categories, while the people on the left are the opposite. So we think people are aging like cars – the whole car gets older, but some parts are starting to wear out faster than the others.

We think this is actionable information because there are clinical markers associated with these ageotypes (like Hb1c). To dig into this a little bit deeper, we looked at the specific markers and how they are changing within the group also with different interventions. 

So why is this a big deal? There’s a ton of stuff people can try to improve their healthspan, but how do you know they are any good and that they are working? We don’t have good markers for that, DNA methylation is probably one of the best ones out there, but it’s still a little bit slow compared to what we could do with those ageotypes I would argue. In principle we could actually try to do it better.

We can also look at our cohort and see how people are aging using other markers people have done – this is Morgan Levine’s PhenoAge. Sure enough we find people that are above the line that seem to be older biologically and under the line that seem to be younger with a good slope.

Obvious question is are we going to scale this? As an academic project, this is a really deep and expensive endeavour. We spun off a company called Q.bio (conflict of interest disclaimer), and they are doing a medical version of this along with whole body MRI. And right away with a few hundred patients they found things that are pretty important. Either by multi-variate evidence, when we see several lines of evidence, which is important because with patterns in different markers we can see it earlier – when you jump but are still within the range, it will normally get ignored, but we would flag that. Or by following the changes longitudinally, which is something almost totally ignored in today’s medical world. It’s not cheap, like $3500, but over time we will get it cheaper. So I hope this is the way to scale it.

We’re very keen on wearables, because they are relatively inexpensive and something we all can do today. They are pretty powerful, some of them take up to 2.5 million measurements a day, and they are measuring all kinds of things depending on the device.

We learned pretty quickly that you can detect quite a lot, for example infectious diseases, I detected a lyme disease from my smartwatch and pulse oximetry data for example. I discovered a drop in blood oxygen and elevation in resting heart rate and it wasn’t on a plane, so it was very clear, so I went and later got diagnosed with lyme. When I then looked in hindsight on my data, there were 4 times when I was clearly ill as measured by CRP, and sure enough each of those times, I had a high resting heart rate as well as skin temperature. So we went on and wrote an algorithm that worked well for heart rate where we can build people’s baseline and then look for the shift from baseline to plot a delta line. And in all cases we could pick up the shift in advance of symptoms.

We were working on scaling and improving these algorithms and then along came COVID. So we asked whether we can detect COVID from smartwatch.

The gold standard for testing right now is PCR, which is too slow and expensive, and thermometer, which is fast but unreliable.

So we wanted to see whether wearables could actually help. We launched a study with >5000 participants and >30 golden data points when we clearly knew they were positive and were wearing their smartwatches, and we knew their diagnosis and symptom dates.

This is our very first participant, where we show in the data that their heart rate jumped up 9 days before their symptom and diagnosis date.

Turns out we can detect 80% of the people (it’s not 100% because it’s hard to get a stable baseline for certain people, those are 20%). But for most people it does work, we can detect it median 4 days prior to symptom onset and 7 days prior to diagnosis, for some cases quite earlier. Other illnesses will show this as well, median 2 days. That makes sense, because COVID has a longer pre-symptomatic period than most other illnesses. However it also picks up other stressors, like psychological, alcohol binges, etc.

So we built an online detection system that takes your baseline and looks for the shift – we have several kinds of algorithms for this. One around resting heart rate, another is heart rate + steps + sleep, and different combinations of metrics. It does work pretty well, you can see these alarms in the retrospective data that are triggered well in advance of the symptom onset.

Last year in November we launched a real time alerting system that’s pretty sensitive and scalable, and in principle can follow millions of people, pull their smartwatch data in real time and ping them if there is a sign of illness.

It does work, we have several thousand people signed up already, this is one of the examples where it sent a red alert 3 days prior to symptoms. We have around an 80% detection rate – at or before the symptoms date.

The other thing that’s pretty good about it is that we can pick up asymptomatic cases. These are people who are getting randomly tested and have a positive signal without symptoms, and we look at their data and sure enough, it’s pretty clear there was a shift before the diagnosis date. This alerting system works for Fitbit and Apple Watch right now.

This isn’t the only kind of data we want to gather, we think we can get other things and not just infectious diseases. We built some machine learning models to try to predict people’s clinical measurements. You can get some prediction value, not clinical grade predictions, but enough to tell you that something might be up.

We’re very interested in measuring glucose. We started putting continuous glucose monitoring devices on normal, pre-diabetic and even some diabetic people. We discovered that some people have good glucose control, some are moderate spikers and some are pretty severe spikers.

We can take the data and start classifying what these patterns look like and into what we call “glucotypes”, where people are either low, moderate, or severe spikers, which is not a good thing.

Then we started testing people on different foods, and found that different people will spike at different foods. Nearly everyone spikes to white rice and cornflakes with milk.

It has been shown that at least part of this can be explained by the microbiome, but there are other factors as well. My methylome shifted both times after I became diabetic, so I presume there are epigenetic effects going on touching the metabolic genes that are most affected. And we started looking at people and classifying them, and there are some people who are severe spikers but healthy by all states of measurement, which surprised me – they were spiking just as badly as diabetics. So we think these measure slightly different things, which relates to the subtyping I mentioned earlier.

So to scale this, we started a company called January.ai (conflict disclaimer), which uses CGM and builds a database that can then show you statistics, recommend food, etc. We’ve done a simple study showing that just wearing a CGM and using this simple app actually improves peoples’ time in range – the amount of spiking that they have.

We’ve discovered that we have healthy people spiking like pre-diabetics or Type 2 diabetics, similar to the previous study.

Here’s shown that people who wear CGM even for 10 days can actually improve their spiking. They are very eye-opening devices once you see what spikes you. Over 90% of healthy people improved their time in range and even over 60% of Type 2 diabetics did.

Academically we’ve built what we call a personal dashboard where we bring in peoples’ wearable, clinical, and other data.

And what we’re building now is a personalized measurement for following people over time – very interesting for things like chronic fatigue, etc. We’re trying to follow personal measurements and correlate good days and bad days for chronic fatigue to see whether we will be able to better manage their health and productivity.

We think this all is going to be very powerful, especially with the ability to share with your physician – we think clinical data and wearables is much more powerful then what can be measured in the doctors’ office, and that wearables should get incorporated into medical care.

This is where I see the future going, people getting their genome sequenced even before birth and combined with personal omics we should be able to better predict risk, diagnose, monitor and treat.

Here’s another interesting study we’re running remotely, we have a lot of them in the middle. This one is about the diet and relationship to cognitive function, especially regarding the role of fiber. If you are interested in that kind of study, let me know, I’d love to have more people join the study.

Here are the folks who have done the work.

And we have also a very large team on the wearables side of things.

Q&A

Have you done measurements for baseline heat shock response elements?

  • I don’t think we did a heat shock response, we did some cryotherapy but haven’t got the data back. Others are doing hyperbaric chambers, so it will be interesting when they come back with data.

 

 

How can we scale the kind of science you discussed esp. data collection (many data types at healthy baseline + tracked longitudinally + extra samples in illness): ie (a) How can we get funding for a larger cohort across multiple institutions, (b) can we separate the data collection from the analysis with all the data open to any researcher who wants to analyze, (c) can we incorporate citizen science where high end people contribute but without complete data?

  • It’s hard to get money for this sort of thing, because people tend to study illness and not health. We’ve struggled to get money, the way we slapped the current funding together is with a lot of philanthropy – that helped a lot. We did get one NIH grant when I packaged it under pre-diabetes, which seemed to help. It seems that there is also a rising wave of awareness about wellness and health, so I’m hopeful that will also lift our boat. There’s a lot of mini national projects, in Sweden for example – not that many make their data open though unfortunately. There is also the 100000 UK genome project which is interesting.
  • As for concrete steps, we’ll obviously keep trying to plug into NIH to give us more money. In the wearables project, we’re going to put the anonymized data public – we’re actually the only group that puts wearable and CGM data public. Amazon gave us $1.3M in credits to build a data ocean to put this data out there, so that’s what we’re doing. We’ll also put some limited omics data out there as well. We love people who put their data including genome sequence open so people can use it. Actually more people have analyzed our data than we have, so the data we have is already out there and people are using it. 

 

 

Q.bio seems to be mostly MRI based — is there anyone integrating all these datasets together?

  • I think it is primarily because of their early days, as they wanted to stay around actionable information. There was a company called Arivale, they were trying to build these networks that were from the research standpoint pretty cool, but the information wasn’t actionable, and the company didn’t survive as a result of that. So we are pretty cognizant of that and need to start with a proper business model. So we’re doing all this research in the lab, but then use the actionable parts in the company. But they want to integrate these datasets, it’s only a matter of time.

 

 

Why doesn’t Q.bio offer the package minus the whole-body MRI? (You aren’t doing the MRI piece to the 109 person cohort in the academic work, right?)

  • They are a new company so they can’t offer too many specialized packages. And even if people would like to replace the MRI with the GRAIL test for example, they are not necessarily replacing each other, rather that together they are going to be more powerful.

 

 

Do we already know which more invasive & expensive biomarkers & measurements we can cleanly and reliable cross-correlate with non-invasive, cheaper ones? (extreme example: Facial aging selfie –> DNA methylome)

  • We’re very interested in that, as you saw in the presentation we are trying to correlate and predict the clinical measurements with the wearable data we can extract. We’re also doing a lot of microsampling and pricking and doing deep measurements there. We’re trying to get in that direction and see what we can learn. I am most interested in smartwatches, because there’s already a huge population and I honestly believe that we could put a smartwatch on every person on the planet, 60% of the planet has a smartphone so if we just pair that with a smartwatch, we can really improve peoples’ health.

 

 

You mentioned epigenetic methylome. Is that data for the 109 people as well as the battery of tests they’ve done open to the public as well?

  • Yes and no, we did it as a pilot on me, whole genome by sulfite sequencing (the whole methylome) with a high temporal frequency. We did it as a pilot to see what we can learn and to see where we will screw up. These samples are very very precious, we have been waiting for the price of the sequencing to drop enough to the whole batch – we have like 2000 samples. We want to do it all but we want to wait for the price to drop. So you can get my data, it’s available and out there, feel free to play with it. For the rest stay tuned, I keep thinking we will do it in 2021 but now I’m guessing 2022 might be more likely. 

 

 

How does the HR algorithm avoid alerts on exercise but pick up disease-associated HR spiking? And could it be used for long COVID diagnostics eg post-COVID POTS can result  in major HR spiking?

  • We subtract 10 minutes of heart rate after you run, we can see when you are running using steps. But walking is probably a better measurement, because it’s like a little bit of stress, and it provides a way to get larger signal from the heart, it sort of amplifies it. So we want to tune the algorithms to pull the stuff out when walking.
  • As for long COVID, we are looking at those data now and we definitely see alerts and other signals going off for several months after the infection. So we’re going through the data now and trying to put it together, I don’t have a complete story yet. But we can definitely detect post COVID signals.
  • Related to this is an interesting paper from United Health group that has shown that something like 4% of people get Type 2 diabetes after viral infection, so that fits very nicely with the post COVID recognition – I think this stuff is very generalizable.

 

 

What’s the willingness of payers (employers, insurance companies, CMS) to pay for prevention-oriented interventions? Some VCs are reluctant to invest in prevention companies if payback period is > 18 months – given patient churn in health plans.

  • This is one of the ways in which our healthcare system is broken. We don’t pay to keep people healthy, we only pay once they get ill. It’s really backwards. How do we solve the problem? Well Q.bio right now is a concierge service, we want to get out there broader, get it cheaper, starting out expensive because of a lot of R&D costs, but as we get the scale and new technology, I hope we can get less expensive and out to the world. So I hope once it is cheap enough, they’ll take it on. In the meantime if we can make at least some parts of this, like the wearable data, show that everybody should be doing this, and if those plans could have that, I think that would help. They tend to roll it in in a different way though. Several groups that we’ve talked to want to do monitoring of at risk populations, or outpatients, and show it works with those folks. So start with at risk groups and concierge and show that it has utility, and then spread it broader. 

 

 

You presented several examples of reversals in different measurements for example in response to lifestyle changes. I wonder whether this applies with aging, for example in the case of ageotypes or maybe wearables, whether you can see it going up and down.

  • We are extremely interested in this. We now have about 7 or 8 years worth of data, and we’re just going through the analysis now. Maybe next time I present that, we didn’t have a clear story yet. Once we have enough data from all the folks to be able to look in a lot more detail, we’ll do that as well, there are a lot of lifestyle interventions people went through. 

 

 

We have funded companies that do self-insured employment. There are two economic benefits – cost savings to your bottomline, when you make employees healthier, and then of course also the productivity increase that also adds to the bottomline. So are there health interventions that have productivity improvements that insurers should pursue if they have those types of plans?

  • I am in a conflict of interest here but a no brainer would be January.ai and Q.bio. Glucose monitoring at least personally helps, since I can put myself to sleep with a slice of pizza, but now I know exactly what foods are going to do what in my body.

 

 

What can this group do to help you further your work?

  • Sign up for as many studies as you want. If you are in the Bay Area and want to join our fiber study, we’d love that. If you are anywhere you can join our fiber and cognition study. You should all be wearing smartwatches, I am wearing four. By all means sign up, even in the CGM studies we have a remote version of that. All kinds of studies out there. You can find that on our website.

 

 

 

Seminar summary by Bolek Kerous.