The Translational Mixer
The Translational Mixer
Episode 5: Nathan Price on scientific wellness and a Mojito
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Nathan Price, currently on leave from the Institute for Systems Biology in Seattle and Chief Scientific Officer at Thorne HealthTech, talks to JC and Andy about his data-driven approach to understanding health and predicting personal trajectories into disease as we age.
03:24 What is scientific wellness?
06:53 Correlates of scientific wellness
15:45. Generating hypotheses
18:40 Multimodal over unimodal data
22:04. Biomarkers and individual disease trajectories
28:00 How to intervene to maintain wellness?
30:01 A new era for supplements?
32:26 Single interventions versus combinations
37:59 Racial background and lifestyle
41:00 Digital twins, trial design and recruitment
43:30 Mocktails and mojitos
The Mojito
10–12 mint leaves and mint sprig
1 Oz simple syrup (50% sugar solution)
2 Oz white rum
0.75 Oz freshly squeezed lime juice.
~1–2 Oz club soda
DIRECTIONS: Place the mint leaves in a shaker tin, add the syrup and gently muddle the leaves 10–12 times. Add the rum and freshly squeezed lime juice. Shake over ice for 15 seconds and double strain over fresh ice into a Collins glass. Top up with the club soda and mix gently. Slap a mint sprig to release the aromatic oils and add it as garnish.
Sources mentioned in the podcast
The supposed Native American 'diabetes gene'. Newman, AS. Peace Rev. 12, 517-524 (2010)
Todd Rose. The End of Average (HarperCollins, 2016).
Can a biologist fix a radio? Lazebnik, Y. Cancer Cell 2, 172-182 (2002) https://www.cell.com/cancer-cell/pdf/S1535-6108(02)00133-2.pdf
Lancet Commission on risk factors for dementia: Livingston, G et al. Lancet 396, 413-446 (2020) https://www.thelancet.com/article/S0140-6736(20)30367-6/fulltext
Precision Medicine Approaches for Developing Combination Therapies for the Treatment and Prevention of Alzheimer's Disease (AD) and AD-Related Dementia, National Institute of Aging, December 4-5, 2023.
Khullar, D How to die in good health. New Yorker (April 15, 2024)
The Mixer music “Pour Me Another” courtesy of Smooth Moves!
03:24 What is scientific wellness?
06:53 Correlates of scientific wellness
15:45. Generating hypotheses
18:40 Multimodal over unimodal data
22:04. Biomarkers and individual disease trajectories
28:00 How to intervene to maintain wellness?
30:01 A new era for supplements?
32:26 Single interventions versus combinations
37:59 Racial background and lifestyle
41:00 Digital twins, trial design and recruitment
43:30 Mocktails and mojitos
Andy Marshall: Welcome, my name is Andy Marshall. This is the mixer. I'm back with my chum and partner in crime Juan Carlos Lopez.
Juan-Carlos Lopez: Andy, it's very nice to be here. Who are we talking to today?
Andy: So we're talking to Nathan Price. Nathan was faculty at the University of Washington and then associate director at the Institute of Systems Biology in Seattle. He also co-founded a company called Arivale with Lee Hood, which focused on the concept of scientific wellness and we'll be hearing a lot more about that in the conversation. But most recently, he moved to Thorne HealthTech, which is a dietary supplements company as their Chief Scientific Officer.
JC: I'm really looking forward to this conversation because health and preserving health is something that we seldom talk about in the podcast. We normally talk to people who are trying to cure disease by identifying new targets, new pathways, new therapeutics. Talking about preserving health is unusual and I like Nathan in particular because his approach is very data-driven. Wellness is something that is very trendy. There are a lot of people out there who claim to have profound insights about this, but they are not as rigorous as Nathan has been with his models and with his digital twins and all the different things we're going to be hearing about.
Andy: Yeah, I think this is what is most exciting about Nathan's approach to thinking about interventions, whether they're lifestyle interventions or dietary supplements. Yeah, as we know, the dietary supplement industry isn't really renowned for its rigor. So I'm really looking forward to talking to Nathan and hearing more about his approach. So let’s get going
JC: Let’s do it.
Andy: So, welcome, Nathan, great to see you on the show. We really wanted to have a chat with you today because we've been following your work over the years. You've been a big proponent of this idea of a data-driven approach to understand the progression of disease and also thinking about how that path leads to disease over time. Over the years, you've focused on how you could take advantage of some of the kind of new technologies RNA and DNA sequencing protein and metabolite profiling deep phenotyping , all of this good stuff, and see how that could be applied to build models to predict that progression to disease. So you've thought more than anybody else in terms of how you put these concepts into to motion and last year you put out a book, the Age of Scientific Wellness, together with Lee Hood, which I think caused quite a stir. So perhaps a place we could start is by asking you to define what you mean by scientific wellness and how you came to think about health and disease in that way.
03:24 What is scientific wellness?
Nathan: Yeah, happy to do it and great to be with you, Andy and JC. The concept of scientific wellness. was basically just the notion that in traditional approaches to health and disease, and this is both in the healthcare system as well as in the research enterprise, we tend to think a lot about late -stage illness. What are your symptoms? You get a diagnosis, you get a drug, and we've set up all of our enterprises, like the National Institutes of Health, right, or a series of different disease institutes, and we focus with them. the context of R01s or different things that we're writing, typically on late-stage disease, it may be an easier p-value, things line up and so forth. So we have this kind of traditional model.
So scientific wellness was the idea that we're actually intervening in disease way too late, and that we've set up the whole enterprise from start to finish around that paradigm, and that we should change that. And so the idea behind scientific wellness is that we wanted to take this same level of rigor that we have applied to the study of disease and apply it to the study of health. So when you start studying healthy populations, you're now thinking more about transitions towards disease, what happens early rather than late, and we'll talk more about that. And so it just starts you down a little bit of a different path.
And in fact, right before the pandemic, I was on a panel with Denis Ausiello, who some people may know, the former chairman, at Harvard Medical School. And he put this in a way that I really liked and so I haven't forgotten it, which is that “healthcare is the only industry that does not study its own gold standard, which is wellness.” And that[s what we're really trying to change.
Andy: Yeah, it's really interesting, I think. And it's also something that we can explore in detail over the course of the conversation is the way in which everything in the research industry enterprise is focused around disease. It's the National Institute of Health , and yet it's organized around disease.
Nathan: And I think that ends up being pretty interesting because health can be amorphous, wellness can be amorphous. We actually debated a lot about even the term wellness. The term wellness kind of had a negative connotation to it and in kind of interesting ways. in the scientific community. But one example of what this means in terms of focus on disease, there was a study that was done on the genetics of a Native American tribe that was in the areas kind of around Seattle. And the paper gets into how they have ‘diabetes genes’. Now, of course, they don't actually have ‘diabetes genes’. They have genes that are really well adapted to their long -time natural way of life that are maladapted to a modern lifestyle, which of course does have the product that they get diabetes at these higher rates.
Well, and I'll say one other thing, which is the way that the genome was really sold to the public to me was really wrong. You know, this is kind of in the early stages, you know, on through many years. That was the notion that because we're so focused on disease, we learn genetic signals related primarily to disease. So the genome then becomes like this sort of crystal ball that's gonna tell you how you're gonna die, right? And people used to say things like this all the time, don't get your genome sequence, you're gonna find out that you can get Alzheimer's, and there's nothing you can do. That's not true, we can get into that later. But they would, you know, would talk about that. So really though, the genome is a book about your life. It's about how your body operates, and you should be able to learn things about how does it operate optimally, not just what are the ways that it gives you higher risk for this, that and the other. So it's a wellness point of view instead of a disease point of view.
06:53 Correlates of scientific wellness
Andy: So could we talk about how you go about this, what the process is? Obviously, you're going to build a model using an initial starting cohort of people and then you're going to build that out. Talk us through Nathan, the types of data that you would collect and where you would get it from. Maybe talk about some of the projects that have set up kind of leading up to now and kind of what's the present status, like who is doing this now and what are the main initiatives where people are gathering the type of data that you need to do this?
Nathan: Yeah, and it's a great question 'cause it's... pretty easy to say something like, well, we should have more of a wellness -centric focus or we should think earlier, and that's not in and of itself so novel. What's really novel is trying to put together, as you'd say, enterprises to dive in and make that real and more technical and have some depth to it. So the way that Lee Hood and I really started about this was to start generating a lot of really dense data on healthy populations. We did this first in a project that we called the Pioneer 100 and we had a company called Arivale, which generated a lot of this same data. So basically that led to a bolus of data of the type I'll mention here in a moment on about 5,000 people. And so this was doing whole genome sequencing in the early days, later on SNP chips. This was about 1,200 to 1 ,400 analytes out of the blood in terms of the metabolites proteins, a bunch of clinical labs that people would do, measures from the gut microbiome. What are the species there? In the early days that was 16S RNA and Arivale later with you know some of the companies I'll talk about like my current company Thorne that turns into metagenomics. We had health coaches so we knew what people, what their goals were. They would check in on a regular basis about what they were doing so a lot of text-based lifestyle information. It's interesting because a lot of of that text -based information is now much more minable because of large-language models (LLMs) and all the advances now.
So you can do that in a much more advanced way. I'm sure we'll get there. What we found is that, so you start putting together all of this dense data on individuals. And when Lee and I first started this project, it was interesting because we got a lot of pushback on this. And a lot of people were really strongly of the opinion that this was a complete waste of time. I was kind of surprised by how vehemently some people felt this because I'm like well if I'm wasting someone's life, I'm wasting my life. I should be upset about this. And the reason for that was it's there's not a case -control design, right? We're not looking at disease versus normal, looking at the differences, and doing the things that we typically do. And by the way, I've done tons of those studies as well. I'm not denigrating that. A lot of those. But that's not what we were doing here. So that leads you into an approach then about, well, what can you actually learn from individuals when you're just monitoring their trajectories? I'll give a few examples here just to get us going.
So one example is that you can take genetic risk as a rubric to start stratifying people. So initially, when we started this, we were thinking about getting to 100 ,000 people right and following their trajectories over time, and we’d watch for the emergence of disease. We, of course, did that, but that's a long process to get at things. And it occurred to me as we were going through this in the early days that we didn't have to wait for that, because you could take these polygenic risk scores that had been developed. In fact, they're radically better now than they were then. We'll talk about that if we want. But basically, you get these polygenic scores, and so you know which people in the population as a group are more likely to transition to many diseases than others. And so the very first thing you can do is just stratify in these healthy population all this dense multiomic data based on low risk to high risk of these diseases and see what's different.
So we did that for 54 different diseases and conditions. And we came up with, it was on the order of 750 different correlations between those diseases and these analytes. Now, this is very... different than what you do in a case control design because you're analyzing people who don't have symptoms, so no one has symptoms here, but it leads to a very different style of analysis. So when I do a typical case control, or any of us do a typical case control design on late -stage disease, with omics data, transcriptome or proteome or metabolome or whatever, you tend to have hundreds of differences in there, maybe thousands. And then you have to sort out three big things. One, what might be the cause of this disease? What are downstream consequences of that cause, the cascade of other events? And three, what are compensatory mechanisms, things that the body is doing to try to make up for that underlying problem? So if you just focus on differences, right, some of them are good, right, some of them are protective, and we've gone into that in some detail. Now, if I'm looking upstream in people that are asymptomatic, I don't find hundreds of differences. I find very small numbers of differences that are statistically significant, but they're more likely to be enriched for things that our early-stage causes or related to that mechanism.
So, for example, when we did this for coronary artery disease, and we were looking at about 400 different proteins, it turns out there's only one whose concentration in the blood is correlated with genetic risk for coronary artery disease: PCSK9. PCSK9 (proprotein convertase subtilisin/kexin type 9) is the only one, of course, anti -PCSK9, the biggest blockbuster drug in that space in the last 10 years. Asthma, similar thing, interleukin (IL)-33, the only protein that is different in individuals who are at high versus low genetic risk in our data. And lo and behold, we look that up, there were four drug companies that were in late stage clinical trials, targeting IL33 as a potential therapeutic for... asthma and on and on.
And so, you know, there were many of these cases. So what it does is that when you start studying people who are basically healthy, but you have these genetic rubrics to differentiate, you get these very sparse, but at least to my eye, very significant signals that I would argue are more likely to be related to the early-stage triggers of a disease than trying to sort it out once there are hundreds or thousands of perturbations downstream. So you basically let you see a movie rather than a late snapshot. That's one example.
So another thing that we found in these data were that because you have the stacks of multiomics, we looked at the effect on statins. So statins are one of the most commonly used drugs. So we had, of course, tons of people that were on them. So we did an evaluation to see what data types seemed to predict the efficacy of statins. And it turns out that the data type that is really predictive of the efficacy of statins is the microbiome. So we took the microbiome, this was joint work with Sean Gibbons, who is this brilliant microbiome researcher at ISB (Institute for Systems Biology, Seattle, WA) one of my favorite collaborators. What we did is we broke this down into four groups, just unbiased clustering, and then we evaluated the effective statins. And what it turned out was that for one of those groups, the amount of lowering of LDL cholesterol from a statin is double what it is for other people. Now there's no genetic signal that's even close to that strong.
Now a side effect of statins is that you get this 9% increase in the rate of diabetes amongst people that take statins. And so we then looked at the microbiome and again it predicted that two of those groups saw significant increases in their diabetes markers but two of them did not. So again the microbiome was predictive not only of efficacy of the statin but also of whether or not you might be at risk for transitioning to diabetes.
The point I want to make here in terms of this different paradigm is if I go down the traditional route and I want to find out the effect of the microbiome on statins, I have to come up with the idea, write it up, write an NIH proposal, get rejected, write it again, get it accepted, run the study, recruit the patients, run the trial, etc. etc. I'm five years in by the time I have a chance to answer that question. When you've run a scientific wellness program where you just are gathering data at this density all the time like we were with the company (Arivale), I can know the basic answer next week. There you go to a poster. Let's look this up. Does this have an effect? Does it not? So that was the thing that was really amazing. Once you start stacking data really deeply in this kind of what we call scientific wellness or data-driven wellness approach, is that as you get more and more of those data, you can go back and ask question after question, after question, after question and write papers and studies.
And it takes a while, right? You're sorting through multiple hypotheses, you're correcting. You know, there's a lot that goes before you would publish it, but you can get yourself to answers very quickly. Generating these large data sets that are very dense to me is just an essential part of biomedicine going forward.
15:45. Generating hypotheses
JC: So how much of the approach is about generating hypotheses as opposed to testing the hypothesis? Because in the example you've given, so I think they are very interesting and very compelling, they do sound like they are hypothesis generating that then you would have to go and test. Is that a fair interpretation of the approach?
Nathan: Yeah, I totally agree with that because you end up spinning off observations at an incredibly fast level and you're exactly right. So if you wanted to say, in the way that we typically do it's like you're gonna stay on one of those problems, go deep and prove this out, that's exactly right. So for example, in the statin example I gave, we find this really compelling signature. It's done on thousands and thousands of people, so it's not a small observation. But the obvious next step is, okay, but let's do a clinical trial and say if we shift the microbiome in the direction from a low -efficacy statin microbiome to a high -efficacy statin microbiome, does that in fact induce the change that we want and then can you turn that into something like a reliable therapeutic? So you're exactly right. So I think the big scientific wellness studies are primarily hypothesis generating.
Another good example of that is we looked at people in Arivale cohort who went on to lose weight versus not. And again, we tested all of the different types of data to see what was predictive. Controlled for BMI (body-mass index), right? BMI is highly predictive. But, you know, is the metabolome predictive once you control for BMI? And at least in our data, it wasn't. There was, again, only one data type that was. And it was the gene content of the microbiome. The correlation was 0.3, right? So you're capturing a fair amount of variants, you know, not radically high, but reasonable. When we looked at that, there were two features that mattered. One was how fast the microbiome, particularly the Bacteroides, were growing. And the second bigger factor was that the gene content suggested that you were more likely to break complex carbohydrates down into short chain fatty acids.
It was easier to lose weight. But if your microbiome was preferentially breaking down complex carbohydrates into simple sugars, harder to lose weight. Now, that's an observational study, right? It comes out, makes some sense. But that would be an observation, right? You wouldn't say like that's proven and tell you go forward and you do exactly like I said You would say all right We got to set up a study I'm gonna start people at ground zero or even just predict on the microbiome and watch it go forward or else that so you're exactly right, It's very hypothesis generating But the other element that I think is so important is now if I'm setting up a study I know that the microbiome looks to be a really important aspect I'm not gonna spend a bunch of money generating the proteome and the metabolome necessarily because that wasn't actually predictive in that sense.
So these really broad dense studies give you a sense of what's a question that's likely to be yes and what are the data types that you actually want. It really lets you, I could GPS in on the study that you wanna do.
18:40 Multimodal over unimodal data
Andy: So Nathan, you've given this example, well, you've given several examples that are really compelling. And yet many those were validated using a different approach. Obviously, polygenic risk scores and genetics is really established. Many drug companies have their own, like Regeneron has the Genetics Institute and, you know, Amgen bought DeCode to generate specifically genomic data. So there's different levels of data that you're putting on the this, but it's kind of obvious that by doing this, you are going to generate signals that you wouldn't see otherwise provided. You can deal with statistical quandaries of large data sets.
But the flip side of that is somebody might say, well, why do we have to spend all of this money on measuring all of these things? A lot of the things that really matter most are in the plasma. Let's measure the plasma proteome. And it's clear that, yes, that's not going to find everything that you're going to find. But the question is, is like, how much can you find by not going the whole hog in the way that you are, like measuring everything and kind of doing these much more kind of restricted approaches, do you think?
Nathan: You're pretty limited, I think. Now, you're gonna find a lot, right, for sure. But, so the weight loss example I gave, right, proteome didn't pop up. When we looked at trying to predict the microbiome diversity, it showed that you could predict the diversity of the microbiome from the metabolome, but again, not from the proteome. The proteome did not capture it. So there's definitely been cases where we find that, you know, one kind of data can carry a much more larger signal.
Now, there are things that you'll definitely find a signal in all data types, like we did a biological age, multi -omic analysis, and yes, you can definitely learn something that's going to relate to age or sex or disease or like any kind of big biological phenomenon that is predicted from the metabolome, that are predicted from the proteome, that are predicted from whatever, and the accuracy is at least in terms of like the relation to age, like how these things take up over age was relatively similar. I think the metabolome was a little bit better. But if you look at the correlation between, for example, a biological age estimate that you get off of a metabolome versus a proteome, they're pretty highly correlated because they're all correlated with age.
But if you look at the delta age, so the difference, whether or not you're biologically older or younger than your actual chronologic age, those are only correlate at 0.2. So if you take epigenetics or metabolites or proteins or some different biological source, even correlated against something as strong as age, the Delta is only correlated at around 0.25. So it's there. It's a signal. It's significant, but they're definitely telling you different aspects of information. The metabolome is much more tightly connected to the microbiome than the proteome is. For disease diagnosis, for what it's worth, we found the proteome to be the most informative of the data types, at least that we had analyzed, you know, the studies that we did. So I tend to think like, if I'm doing a big study and I want to try to develop biomarkers for disease, proteins seem to be probably the best, you know, source of those. Not surprisingly, we use a ton of them clinically.
22:04. Biomarkers and individual disease trajectories
But some of these approaches, you know, and this gets us into a whole other area. But I kind of want to move beyond the concept of biomarkers entirely, at least in principle. And so this is where we get into some of the things that we talk about, like digital twins, or a project that I'm just initiating now that I call the N-of-1 analyzer. What I mean by that is, so a lot of the papers that we've published have been big statistical analyses on lots of multiomic data, right? And there's, there's reasons for that, right? We find out these patterns, we find out what's interesting, we get hypothesis leads, et cetera, et cetera. Well, what I actually wanna be able to do is to take an individual patient and given a lot of this kind of information, say something about their health, their trajectory, something specific, actionable, can we solve their problem?
So in terms of this thing that I call the N-of-1 analyzer, the idea there is, that we're able to ask questions about a person's individual trajectory. So, for example, what can be done in this kind of case is that you take, we build a background distribution based on all this data that we have, right? So what's kind of typical variation in the metabolome or the proteome or the microbiome or whatever, right? And you have to define that. And there's all kinds of things that you can go down and define that, right? is healthy has to be defined, what's the right age range, et cetera. We could go down a rabbit hole on that. But then what I do is I take an individual's trajectory and I compare it against this background and I make an assessment of how unlikely it is to see the values that I'm seeing. Now you can do that on multiple fronts. I can do that at an individual analyte level. And this is by the way, is what we do in all of medicine. This LDL cholesterol is really high. We don't expect that to be that high in a normal population.
So it starts at a pretty basic kind of thing. And you can just sum up a vector, right, of just ticking what's unusual in this person. I can do that at the level of relationships, meaning this is like building a Bayesian net, right, where you say glucose is within range, but it's really strange that, you know, glucose and hemoglobin A1c are all uncoupled. They usually predict each other, and now they don't. Or, in other cases, things that aren't really coupled and they now are coupled, so you can look at relationships. You can do it based on a priori -defined pathways. You can do it on dynamics, meaning that if you have time series that something jumped, you're still in range, but we have examples I could walk you through of people where their CEACAM5 levels, just jump dramatically, still in range, they were low range to high range, and that was a signal that they would go on to develop in this case very sadly pancreatic cancer and die two years later; every single person in the Arivale cohort—there weren't very many that had metastatic cancer— but every single one of them exhibited a jump in CEACAM5 18 months to two years in advance. So the idea of the n-of-one analyzer though as you start to build all of this information about what's abnormal in that one person relative to these large statistical models that we've built a background. Then what you can do is you actually do is a mechanistic reasoning. So this is where AI comes in an interesting way. You do mechanistic reasoning on the kinds of things that you see that are different in order to make an assessment of whether or not they have had a transition away from health and then what they might be moving towards or have.
Now the point that I want to make though is is that the traditional approach to diagnostics or we develop biomarkers is all based on an assumption, which is that when a person transitions towards disease, they have to go through a similar disease process or at least a bottleneck somewhere in that process. So that this molecule is always high or this panel is or this signature or however we define it, right, that has to always happen. But now in the context, context of something like the N of 1 analyzer, that does not have to be true. This lets you get towards actual personalized medicine because you're making an analysis on what you understand about the biology of the things that individually you see in every person. This is greatly accelerated by things like LLMs because you can use those to harness information out of tons of papers that maybe haven't been applied before and generate hypotheses. So this is a very AI -driven approach, but that gets us to something that is pretty radically different than what we have now, and it is an analysis of your trajectory. There was a really interesting statistic that came out, at least that I saw a week or so ago, just out of the sports world. So LeBron James, a very famous basketball player, he's had a 20-year career, one of the longest careers ever, and his average stat line is something like 28 points, 7 rebounds and 7 assists.
And so the question was, how often, over the 20 years, has he had that stat line? And the answer is, he's never had it. He's never had the stat line that is his average. And so the point of that, and there's a great book on this called, I think it's called The End of Average, I think, people radically, when they don't understand high -dimensional space and so forth, radically overestimate the degree to which you're ever average on some multivariate analysis. You never are, never. And so the whole notion of this, like the n of 1 Analyzer and some of these things that we want to build, is that you are on your own trajectory. And yes, you learn a lot from all the others. There are obviously commonalities. That's a lot of what the science is built on. But there is this also aspect of this n of 1 trajectory that you're on. So I just think that this is an approach, I think it's really at the frontier of how we look at these personal trajectories, how we go from health into disease, and especially how do we reverse it? You don't want to wait till you have some horrible symptoms and you slightly mitigate them, but we actually want to know is what's going in our bodies decades in advance, and if we understood, could we reverse, can we push that off in time, what are the limits of that and so forth?
28:00 How to intervene to maintain wellness?
JC: So this concept of the N of 1 analyzer I think it's very interesting. From the different things that you will find in each individual this sudden jumps or this sudden deviations and what defines the trajectory of the different people, I wonder how much of that will result in actionable phenotypes if you wish?
Nathan: Yeah so the goal of all this is definitely to be actionable. So some of these things I think we probably will have an ability to action on. So this again is where we'll leverage as much as possible this vast biomedical literature, which again is so much more accessible now with the ability to pull out of it with LLMs and so forth trained like exclusively on PubMed or things of that nature. So that's one.
And the other thing to remember about scientific wellness approaches is that you've got to do interventions that are very safe, because especially if they're not manifesting symptoms yet, right? There's a reason that we give cytotoxic chemotherapy to late stage cancer patients because they're going to die, because otherwise it'd be a horrible thing to give to anybody, right? So this is where we get back into, you know, a lot of what I've done and one of the reasons that, you know, I moved to Thorne some years ago, which is, you know, one of the top quality, if not the top quality, natural product manufacturer in the United States. So a lot of natural products, I think, are gonna come into this as well. So if we know things like lipid levels are high or glucose or something, things like berberine have been shown over many years to have an effect. Or if it's looking at the brain, dietary choline is important. We've done some deep simulations of metabolism in brains with a partner, you know, and you find out that phosphatidylcholine appears to become rate limiting to generating energy under low oxygen conditions. That's an important factor so you could do that.
So I think a lot of the scientific wellness interventions will be a mapping, and there's already big databases of a lot of this stuff, of a mapping of lots of these natural products and targets that they hit.
30:01 A new era for supplements?
Andy: You're making a really important point, Nathan, about this, certainly during our initial conversations about this concept, which I had completely missed, which is... is, you know, obviously with a patient who is not yet symptomatic, it may be years before disease actually becomes clinically diagnosed. But you can you can see that you have this signal that gives you an idea that risk is heightened. So then most drugs, you know, you have this risk-benefit ratio, yeah? Whereas many of the supplements that are out there, there's all kinds of issues (with quality) that you're very much aware of. But the fact is, is many of the agents have been used and have been around for hundreds of years. They're relatively safe. They're having a very, very small effect compared with a drug.
Nathan: Typically, that's true.
Andy: And yet, if you implement that therapy earlier, you can change the trajectory.
Nathan: That's exactly right. And so we've built these digital twin simulations, particularly for brain health, right? Can't do it for the whole body, but we've done a lot in brain health. And one of the things that turns out to be really interesting is that even when you simulate the effect of a supplement or a vitamin or something like this and you simulate it in a population, especially when you're focused on one, you find that the expected effect size that it might be there for some people, but it's relatively small. And sometimes you can't find it even.
But when we do the simulations where we personalize and you do combinations of things that are specific to the individual where you're doing a different thing, person to person, at least in the digital twin simulations, the predicted benefit goes from very small to actually being quite large, like significant. Like in the case of Alzheimer's, it looks like probabable delay of years in terms of onset that you can do if you're doing things consistently over your life.
And we have deep mechanisms behind that, and we get into all that. But the point I want to make right now is that combinations and consistency make a huge difference. And one of the things I love about these digital twin simulations is it lets you run an analysis of what does a lifetime look like under all kinds of different counterfactuals, which you can't do in your own life, right? We can run these large studies and it's one of the reasons I'm pushing within this community very strongly for a new paradigm for how we do prevention clinical trials.
32:26 Single interventions versus combinations
And so the new paradigm I'm really trying to push there is I think that this focus on a single compound intervention is really not a good way to go about it because the effect size of a single intervention tends to be small. You have to get incredibly lucky because it has to be like the final piece of the puzzle. If you have a clock that has three broken gears, let's say, if you take one of those gears out and you put in a new gear, you'll say, "Ah, it wasn't that." You do the same each of them one at a time, you'll find out that none of them work, so there's no hope. There's a brilliant paper on this from some years ago that I just loved. It's a long time, it's like 20 years ago when I was in graduate school, maybe you read it, it's a classic amongst engineers called Kenabalu... "Can a biologist fix a radio?" And the answer is no. But it was great because you have this diagram, right, a diagram of how this system functions. If you apply it like the tools we have in biology, right, you knock out some pieces, right, you knock out a transistor and you find out, oh, you know, a whole bunch of functionality goes away. So you give it a name that's related to that function, and it's a very important component, right, a most important component, right, and you knock these things out. out, but you can't, it was kind of an opus towards systems biology, but it's basically the notion, how do we pull these things together?
And so when we do these in these digital-twin simulations, it's kind of the same thing. Like in Alzheimer's, a good example of this is vitamin D. So if you look in observational studies, elderly people who are low in vitamin D get Alzheimer's at quadruple the rate of elderly people who have high vitamin D. There's all potential confounders and things like that, but there is actually a known mechanism. So when we build out with our partners this simulation of the brain, one of the things that vitamin D does, this is just known from the literature, is that the presence of this reverse cholesterol transport pathway that's really important to energy generation in the brain, we think that's really central to the onset of Alzheimer's for a variety of reasons.
It's regulated over decades, but it's noisy. The presence of vitamin D tightens the consistency of that pathway that's known in papers. So in the digital twin models, when we simulate, well, what happens if we simulate a person's life over decades with this attenuation and you calibrate it to the known level with vitamin D? Well, sure enough, in the predictions, it says, oh, these people are going to get Alzheimer's at much lower rates and the projection shifts. And it shifts, shifts very similarly to what we see in the actual population. So that's interesting. Again, hypothesis generator is not conclusive, but it's pretty interesting. Now, vitamin D is not included as part of the recommendations for the delay of the onset of dementia in the latest Lancet commission, which did a really important paper in terms of what are all the things you can do to prevent dementia, but vitamin D is not included. Now, why? Now, the reason it's not, which will seem, at least on its face, very reasonable is because there's been four randomized clinical trials that all failed. So the conclusion was, quite naturally, correlation, not causation. Okay, maybe true, but I'm going to push back on that a little bit because when we took the digital twin model and we simulated those exact clinical trials (i .e., our simulated patients were of the same age, getting the treatment that was there) these trials were all quite short, right? So over a certain time period, the longest was three years. And so when we did that, and we did the simulation with the digital twin, which again gave us the lifetime risk, right? It was calibrated to that lifetime risk that I mentioned of that quadrupling and old age. When we did that, the digital twin simulations came back as there would not be expected to be any difference in a single interventional trial over such a short period of time.
So in my mind, vitamin D has not actually been shown to not be causal by the RCTs. It's never actually really been tested. And the digital twins let us see that. Now, the flip side is that even if I simulate then a clinical trial on single intervention with vitamin D, I find that I need a long period of time in order to start seeing those disease trajectories. I can shorten it somewhat by targeting it on a certain parts of the curve, because again you get this vision into what does the whole of the trajectory of our life look like. So I can simulate the trial in advance. Yes, that's hypothesis generating, but wow, I would rather do that in advance before I start spending the years and millions of dollars I need to do the trial. But what it also shows is that if we don't run a clinical trial on vitamin D, if we instead run a clinical trial on personalized interventions guided by the digital twin, that we would expect to be able to center a trial. If we pick the right ages and we center it on conversion into mild cognitive impairment, you should be able, by personalized intervention combinations, to get a large effect size that would let us run a trial on, say, 150 people for a year and a half, you know, maybe two years at the outside to see it. Which is way better than needing to run a 10 ,000 person trial for a decade. That's the thing that I'm really pushing on a lot right now, which is I think we should move away as much as possible from these single compound prevention trials because they take too long and especially since prevention has to be safe so it tends to be natural products. The economic return on proving that vitamin D is helpful is like nothing right because because vitamin D is a dime a dozen and it's sold by tons of companies. There's no economic return to running the kind of trial that you would need, right? There's no drug model, right? But if you do these things in combinations, then you can get much bigger effect sizes, much shorter periods. And if you're guiding them with digital twins, so you're recruiting it exactly the right timeframe, right? When, when their most action is happening, you now have a chance.
37:59 Racial background and lifestyle
JC: Can you comment on the effect of racial differences that you may see with this approach? And also, I also wonder about lifestyle differences, right? For example, being raised on the Western diet versus Mediterranean diet. Are you stratifying already when you acquire the data or is the model powerful enough that these will actually flow to the surface on its own?
Nathan: It's a great question. So I'm going to answer in a couple different ways. So first, obviously, what we're mostly focused on is being able to incorporate genetics and genetic signals as strongly as possible. That accounts for more than race, right? That tells you what your genetics are, which is better. Race is a weak proxy for what you get at the genetic level. But it does make an effect. So anyway, just starting from that—so the biggest thing you want to know is all those different factors.
Now backing up and saying what we can say in a course way by just splitting by race. Well, I'll give an example. So I was one of the three invited keynote speakers for the Combinatorial Therapeutics and Alzheimer's meeting at the NIH a couple months ago. There were three speakers. So the first (Jeff Cummings, University of Nevada, Las Vegas) was a really professional well-done pharma talk. Went through all the pipelines where we're at. But it was hard not to be struck by how expensive the drugs are and how small the effect size is. You know, you're talking about maybe a few months of slower decline with pretty severe side effects. The second talk (Laura Baker, Wake Forest University) was on lifestyle intervention and it was hard not to be struck by how massive those effects are in terms of years of prevention, years of benefit at very low cost and there it did get into some of these differences by ethnicity. So it turns out that by lifestyle interventions that the estimate (of disease responsive to lifestyle intervention) now is highest for the Hispanic population. And it turns out that it's estimated that 56 % of Alzheimer's disease could be prevented via lifestyle intervention in that population. They're particularly responsive to it. And that if you look at other populations they were lower than that. We know that APOE4 (ε4 allele of apolipoprotein E gene) gives an elevated risk for Alzheimer's disease. But that degree of which it is coupled is different in different populations. So if you look at four different populations, so Hispanics, again, the relative risk from ApoE4 is small. Then you have the risk in white Americans, the risk in black Americans, and then you have the risk in Japanese people. And the risk in the Japanese is really pronounced. It's much, much higher. And so there's this difference based on the background genetics. Again, coarsely, with race, you do get differences in terms of how important it would be that your ApoE4 status would matter. If you're Hispanic, it's less likely that your ApoE4 status is an alarm bell the way it is if you're Japanese, for example. And there's lots and lots of examples of that. And ultimately, the idea of building digital twins and n-of-one analyzers is that you're accounting for all that, but you're accounting for that plus a lot more, right?
41:00 Digital twins, trial design and recruitment
Andy: This is so fascinating, the different avenues that this opens up. If you can build these digital twin models in different diseases, Alzheimer's is a great case here because there's so much literature there, so much known about mechanisms, you have all of the kinetic data, other diseases, perhaps not so much, maybe the model's not going to be as great, but... could you use a digital twin model to tell you how do you design your trial, thinking about endpoints? Which are the right patients to recruit into the trial that will give you the greatest chance that you're going to get the answer that you want? I mean, this sounds like music, I am sure, to most drug developers' ears. So, in terms of clinical trial design, what's your thought about how you could use this?
Nathan: Yeah, I think it's an incredible resource for drug developers. So one aspect, as you mentioned, which is that you can simulate the clinical trial in advance, specifying what ages do I wanna go, what's the clinical endpoint, right? Do I wanna look at mild cognitive impairment? Do I wanna look at Alzheimer's diagnosis? How much of a longer trial do I need to get to Alzheimer's diagnosis than I need to do MCI (mild cognitive impairment)? Do I want to start with people in their 60s or their 70s? Do I want to recruit different ages based on their genetics? Do I want different ages for my ApoE3/(ApoE)3 people as opposed to my (ApoE)3 /(ApoE)4 versus my (ApoE)4 /(ApoE)4?
Another really big aspect of this, and I talked about this at the NIH even a couple years ago, which was what stage is my patient at? Yeah, you can actually use these. these models to try to control for where in the disease process you actually think your people are and on and on. So you could do like really remarkable analyses for clinical trials. So I think it just gets us to a whole different level. I'm a big believer that if you come in from this kind of systems biology approach and try to understand it at some level, and you can only get the beginnings of it, right? That the body is so complicated, we're always surprised, there's a lot of empirical evidence, but I think you can get your framing. If you get your framing better though, you've got a much better chance.
JC: I think it's a very interesting approach that has been a long time coming, so I look forward to the results that you will see and to continue seeing the new hypothesis that your approach generates.
Nathan: Yeah, I appreciate that.
43:30 Mocktails and mojitos
JC: So as we told you before, the podcast being the Mixer, we combine our interests in science and our interests in drinks, but we know that you don't partake on drinks. So is that something that came out of your analysis of risk factors and your analysis of wellness, or is that just something that has always been the case? Or do you happen to have like a favorite non -alcoholic drink that you can share with us?
Nathan: I'm not a drinker. I just, I grew up in a family that didn't drink. You know, I grew up in Utah in a Mormon family. I'm not religious at this point in my life, but I was raised in that. So I didn't drink as growing up. I don't not drink now for that reason, but now that I'm studying longevity and health and science and things like that, I guess I didn't see a point to starting to drink. I don't know. if I'll do that at some point or not. I'm certainly out with people to drink all the time. I don't have a favorite mocktail at this point, I guess. I had one the other day out in Marin County that was fabulous, but I don't know what the name of it was. They just said this one is really good and gave it to me and I agreed.
Andy: That's basically JC's view of mocktails. He stops listening when somebody starts talking about them.
JC: That's right. Well, thanks very much for joining us, Nathan. This has been very interesting and stimulating.
Andy: Thanks a lot, Nathan. It's been great.
Nathan: Great to talk to you both. Really appreciate it.
JC: Well, Andy, this was a very nice conversation because you and I have had similar conversations in the past about how valuable it would be to have these longitudinal data, these very deep phenotyping and genotyping of people so that we can prevent disease as opposed to cure disease and the fact that Nathan is doing it now to me is very encouraging; it's a step in the right direction. Its very interesting because very recently I read in the New Yorker these statistics that after 65 years of age you can only aspire to one year of healthy lifespan.
Andy: That's a scary statistic!
JC: It's very scary but I think that by pursuing the line of work that Nathan is doing we maximise our chances at living a longer health span. And I think that that's the goal that he's set out to do. So I'm all behind his approach.
Andy: You’re all in!
JC: I'm completely all in. And I'm sure you are too. And you know, this is very interesting too, because as you know, a lot of people in this space of health span are either appealing to wealthy clientele or they are selling snake oil right? So to me, that actually is refreshing about Nathan's approach, that he is focusing on democratizing this health-span concept and bringing a rigor that is not common in this field.
Andy: So that leads to this question of, who is the customer, yeah? Is it the high net worth individual? Is it somebody like us in our 50s, 60s or 70s? And the thought is that perhaps with many of these conditions you need to start earlier, yeah? And can you convince somebody in their 20s or 30s to take up the challenge that Nathan's kind of giving them?
JC: Yeah, that's a pertinent question. We don't know, but the first step is to actually get the data. And the more robust the data is, the uptake will increase. And I think that the fact that he is trying to appeal to in the general population as opposed to a select group of people, I think that's quite welcome. And I really hope that his approach is successful.
Now, let me mention one thing. And I was very disappointed about the fact that Nathan, because of all his healthy lifestyle, didn't bring a cocktail to the table. So I was hoping that you would share with the audience your favorite drink.
Andy: So my favorite drink is a mojito.
JC: Very nice, Andy. The mojito, it's not my favorite drink, but I've made it many times. And the mojito is an unusual drink in the sense that the mint that you put in the mojito is not just the garnish right you first need to muddle some leaves of mint to add to the flavor of the drink and as you may know, Andy, if you muddle too much you don't only release the aromatic oils that should incorporate into the drink but you start releasing chlorophyll just a little bit, releasing polyphenol oxidases, and they turn your drink a little bit too bitter and that doesn't go well with this drink. So you need to muddle firmly but not too much right. So you need to experiment a little so that you really get the right amount of muddling so that you get a very tasty mojito and the next time you come to the house I promise that I'll show you the number of times that you need to muddle to achieve perfection in that drink. And in fact, we're going to share this with our audience down in the description below.
Andy: Perfect. Well, thanks to Nathan for really stimulating and interesting conversation. Thanks to you JC for the intel on Mojitos. And for muddling through another Mixer podcast and we look forward to our next guest. Cheers JC!
JC: Cheers Andy!
Lazebnik, Y. Cancer Cell 2, 172-182 (2002) https://www.cell.com/cancer-cell/pdf/S1535-6108(02)00133-2.pdf
Livingston, G et al. Lancet 396, 413-446 (2020) https://www.thelancet.com/article/S0140-6736(20)30367-6/fulltext