# Hacking Aging using Big Data and AI - Peter Fedichev

- Channel: [Edge City](https://streameth.org/edge_city)
- Date: 2024-10-21
- Duration: 33:04
- Watch: https://streameth.org/watch/6716055d8f864ede03971364
- Download: https://vod-cdn.lp-playback.studio/raw/jxf4iblf6wlsyor6526t4tcmtmqa/catalyst-vod-com/hls/5008r18zb6io6pcg/720p0.mp4

## Transcript

big data panel or like blog. We will start with Peter from Jira AI and the stage is yours. Okay, thank you very much for having me. So I'll tell you about what we do at Jira, Jira AI, a company, a startup that is trying to sort out aging from, let's say, kind of engineering and physical science approach. So first, before I start, I'd really like to point this out all the time, is that we live in an interesting time, the time of what is called demographic transition. I think this is a technological singularity in disguise, because what happened is that people stopped producing new people, as you know. We're not getting younger at the same time, which means that we will have a lot more diseases, a lot less death, and the population will probably stabilize. And whenever such things happen, I mean, this is really a phase transition. We're just living at the point of the phase transition, and we're observing it from the inside. And if you look at a physics textbook, you would know that at the phase transition point, things that go up exponentially are those which represent the future order, the order parameter of the future phase. And I think things like colonization of space, developing of AGI, and solving aging and nature-related diseases are the things that define this future phase. So the goal of this slide is to show that it's not like we should be discussing should we be living longer, should we find a drug against aging. Everything has been already decided on a super organism level, let's say. The only thing that is in our hand is how to make this transition smooth for us. I mean, on the other side of the phase transition, people will live long. The problem is how long and how nice we will be living at our generation and the next generation. The goal here is to avoid unpleasant metastable states on the way to the brilliant future. So if aging is a problem, then of course we should look at what aging is. And reflecting to the definition that Jan was giving today and yesterday, I mean, when I came to the field, I really liked the definition that aging is the exponential acceleration of mortality. And this is pretty good because counting dead bodies is very objective, right? You can be either dead or alive. This is a very objective definition of aging. And from this, aging looks really, I mean, cool. On the left, you see the actual chances to die for a human being as a function of age. Humans are actually exceptional species. I mean, I don't know any other species that has such a range of acceleration of mortality as we do. We are indeed exceptional. Wow. At the same time, there are mammals that do not age. And this is naked mole rat. I mean, this is really the guy who brought me into the field. I physically met them. And somebody told me that these guys do not age. And you can see this is a paper that was sponsored by Google. The chances to die for this animal is a function of age, and you can see that the error bars are going up, but the curve doesn't go up. So these guys are technically not aging. So this means that I think if somebody is thinking about radical life extension, we should really ask these guys or girls if they are thinking about trying to mimic this phenotype or not, because anything that is less radical than this, how to make that guy to that guy, everything that is less radical than this will not solve aging. So the question is, can we actually understand how those guys become these guys, and can we actually engineer this in humans? And solely for intimidating purposes, I will just try to guide you what kind of reasonings and models can actually bring you to the understanding of this phenomena. So at first, and this was my first paper that I published on aging, first we started solving model problems with the idea to understand. I mean, it's a very kind of simple idea. If you have exponential acceleration of mortality in humans, so most probably something is exponentially failing in you. And what is the most famous exponential failure example in nature? A nuclear explosion. So you have one neutron, it goes into a nuclei, hits a few more neutrons, and if you have neutron simplification rate larger than one, so if one neutron produces more neutrons, you explode, right? If you start capturing these neutrons, if you reduce this neutron simplification rate, then you get a nuclear reactor. Neutrons are always produced, but the reactor is not exploding. So our idea was that probably living creatures can be separated by their abilities to amplify noise. For example, the idea was, what would happen if I make a molecular error in a cell? Would this molecular error create more molecular errors, and then they will get amplified and the cell would explode? Or would this error be removed quick enough so that no errors will be produced, and then probably this thing would live long. And this is actually what we have solved. I mean, first we looked at spherical cell in a vacuum, as physicists always do, right? So we built a model of a simple cell, which was built of genes and molecules that are expressed according to the programs in the genes. And we started looking how errors in proteins and in genes are affecting each other and get amplifying each other. And what we found is that if we look at all kinds of animals that are, this is called the phase diagram here. So this is your protein error. This is your protein repair rate. This is your genome repair rate. We found that even in such simple model, there is a phase transition. So if your repair systems are very good, so if you're living here, so if your protein repair rate is very large and your genome repair rate is very large, you are somewhere here, then errors don't get amplified. You are a nuclear reactor. You breed errors, but the average amount of errors does not grow with time. And if you are here, so if your repair rates are small, then errors amplify errors and you exponentially explode. And this separatrix or this phase transition line is probably the line that is separating humans from naked mole rats, right? And then what happens is that there are costs, evolutional costs, energetic costs that are associated with good protein repair and good genome repair, which means that to sit here, you have to expend a lot of energy. And what happens is that the evolution, evolution doesn't like when people waste energy, meaning that all animals are actually forced to this boundary, which means that all living creatures, just evolutionally forced to operate close to this boundary, which is, by the way, called in a fancy way self-organized criticality. And some lucky guys are stable, like at Naked Moral Reds, and some unlucky guys are unstable. So, I mean, with that, we formed a company. I mean, that now sounds funny, but that was really the idea. I mean, look, if everyone is already at the phase transition light, maybe with some therapeutics we can move stable to unstable and the other way around. transition light, maybe with some therapeutics we can move stable to unstable and the other way around. And actually, this kind of understanding brings you to this idea of regulatory forces, like Jan has been explaining today. The idea is that if you have any quantity that looks like a number of errors, depending on the character of the regulatory forces, you could end up in a situation where once you have errors, the regulatory forces remove errors, and then you are stable, and you are fluctuating in a kind of potential basin. Or it could be that your organism is built in such a way that errors produce more errors, and you get destabilized and die. So this would correspond to normal aging when you don't have the regulatory forces that bring you back. And this would correspond to probably inexorable senescence, like no aging. I'd like to stress here is that the animals that are stable, that actually they still die. Because on a bad day, all fluctuations can pile up in the wrong way, and you will go off the stability basin and start disintegrating and die. But the probability of such event does not depend on age. So that was our kind of initial understanding, that probably by finding ways to describe regulatory forces from the data, you can actually understand who is stable, who is not, and which molecular mechanisms are controlling stability. So here is where the modern machine learning comes in, right? So the question is, can we actually identify quantities, the parameters of our models from the data? And the answer is, of course, yes, if you have what is called longitudinal data, when you have many measurements from the same guy. So what we're doing here, we are bringing physics to biology, because physics is a science about trajectories. So if you have trajectories, you can use machine learning in order to understand trajectories in a funny way. So if you have many points measured from the same animal, like many, many medical parameters, you can try to use neural networks in order to compress them into a very few important variables. And then you can try to use these variables to predict the future, and you train this model to predict the future in the same guy and across many, many, many species. And if you ask the system to build a model having just one important feature that would have a meaning of the total amount of errors in the system, then this feature would go up exponentially with age in mice. The exponent would be exactly the same as the mortality rate doubling exponent, so that would most probably relate it to death. And what I like in this graph most is that the red dots are the animals of all ages that are scheduled to be killed in the lab because they are too sick, right? Because in the lab, in aging sciences, you fight the ethical committee to keep animals alive, right? Because most of the time, the ethical committee wants these animals dead. So interestingly enough, from this graph, you can see that mice literally die of reaching the ultimate biological age, right? So this biological age level is incompatible with life. Okay, so nice, you have exponent, which means that these animals amplify errors, they cannot remove errors sufficiently quickly, which means that they are unstable. And unstable on one level is very bad, because the amount of errors is going up exponentially. But on the other level, instability means that there is no equilibrium point, which means that if you have the amount of errors going up exponentially and if you get a treatment that reduces the amount of errors in this system, then once the treatment is stopped, the animals should age at the same rate but from a lower base, which means that for unstable animals, you can rejuvenate. And that's exactly what happens in experiments. These are our experiments. We have two groups of animals here before the treatment. Here we have shock treatment with infinite dose of a drug that reduces biological age. And then you can see that the treated group is always below the control group. This is a proof of instability, right? So, which means that if some animals are biological age younger, they will keep being biological age younger till the end of the experiment. Obviously, these guys are also living longer. You can do infinitely many interesting things with these kinds of drugs. So, for example, this is the fraction of animals alive. This is our drug. And this is the time since the beginning of the experiment, very old mice. This is the preview of the future, because these drugs are now getting into clinical trials. So what happens is that in the control group, you can see that the animals are dying from natural causes, right? Because the time is going on and the animals are dying. These are 100 weeks old mice. They are very old. Then if we keep treating them with rapamycin, you can see that they are living longer, this is known, and if we hit it with a single dose of our drug, because remember we are bringing trajectories to biology, the idea is that we understand the time, the dynamics of the response, we hit them with the drug only once and we see mortality delay, and lots of interesting things like the reduction of the number of senescent cells, blah, blah, blah, functional measures and everything else. So if this thing is not recorded, I mean if this thing is not going online, this thing is being in-licensed by a major pharma now to build a drug against aging. So does it mean that we have a drug against aging in humans? And the funny thing, that is not. I mean, really. We extend life in mice, we rejuvenate a mouse, but I'm telling you now that this will not have a major effect against aging in humans. Why? Because aging in humans is way more different from aging in mice. So, first, remember that in mice, in humans, we have exponential acceleration of mortality, but if you look at any interesting functional, physiological, medically relevant molecular, whatever feature of aging in humans, these are not exponents. These are, for example, incidences of top 20 diseases in humans. None of them is an exponent, right? Interesting. I mean, we're exponentially dying. Our chances to die goes up exponentially. But our chances to get age-related diseases is not an exponent. Now, VO2max. I think most of you know what is that. I mean, this is your ability to actually generate energy, to burn oxygen. And this is how it looks like as a function of age. I mean, this is the number one of the most scary pictures from my presentation now. I mean, look how this thing is going down with age. By the way, if you extrapolate it, it would hit zero at about 120 years old, which means that your capacity to burn fuel will be essentially zero at 120. Look at this. I mean, this is your ability to run at, I mean, run six miles per hour. This is your ability to climb stairs, steep hill stairs, right? I mean, at this point, you will not be able to sleep. So this is not exponent in any way. This is just a trivial linear dependence. I mean, if you're an athlete, it's here. If you're not an athlete, it's here. But I mean, you will be dead at 120 years, no matter what. And by the way, this curve depends a little bit on the chronic diseases, but within this range. Now, this is your cognitive abilities. So this is the battery of cognitive tests. And as you can see, the tests that are related to your ability to speak, which is called crystallized intelligence, they are more or less stable with age. But the fluent intelligence, the battery of tests that is related to your problem solving, is clearly going down and also hits zero at about 120 years old. And by the way, this thing does not depend much on the diseases, because when neurodegeneration sets in, I mean, these things just go to zero. Interestingly enough, that your ability to speak your language model survives aging and your IQ does not. So beware of people who speak well. They sometimes don't understand what they are talking about. So this is aging in humans. So aging in humans is linear, right? I mean, functional measures are degrading linearly, but still we have exponentially accelerating mortality. How can this happen? Even more interesting, that quite some of age-related features in humans are reversible. This is almost politically incorrect to tell, so don't record this as well. I'm kidding. But the effects of smoking are reversible. You can quit smoking as many times as you want before you get a major disease. So these are biomarkers of aging in people who don't smoke, never smoked, smoke, and quit smoking. And you can see that these biomarkers are significantly up in people who smoke relative to those people who don't smoke. But once you quit smoking, they go back to the norm. The ability, remember my example with mice, shock treatment, like smoking for a month, and then the mice are biologically age older till the end of life. In humans, once you stop smoking, you go back. What does it mean? That means that humans are stable. That means that humans recover from perturbations. And you know that. If you go to a gym and start fasting, your body weight would go down. But once you stop doing that, it will go up again at exactly the same level. Right? So humans are stable. Actually, we measured how quickly humans go back to the norm after the perturbations. This is the recovery rate. And you can see that the recovery rate is more or less linearly going down and hits zero at about 120 years old. So at 120 years old, you are not able to recover. I mean, you will tell me, I mean, what does it have to medicine and things like that? This is the most inhuman way to measure recovery rate in humans. This is your inverse hospital stay as a function of age. People get into hospitals in expensive Western systems only when they have big damage. And their ability to recover, as you can see, the recovery rate in those hospitals is going down linearly with age and hits zero at about 120 years old. So interestingly enough, there is no exponential acceleration of any features in humans. Aging in humans is something else. It's a linear decline in our ability to recover. This is totally different from what we see in humans. Aging in humans is something else. It's a linear decline in our ability to recover. This is totally different from what we see in mice. And if we, as Jan has already shown today, if we go to molecular level features like transcriptomes, DNA mutilation, and so on, if we cluster them together to see what kind of time dependencies are there, in any signal, we will find one feature that is linear with h and its variance is going up linearly with h. This is the work we've done with Kirill, who is here. He will answer the questions. And there are lots of features that have kind of hyperbolic dependence. And their variance is actually going up hyperbolically. And the inverse variance is 0, meaning that their variance is infinite at about 120 years old. So what happens in humans is that there is a bunch of things that go up linearly. And remember, VO2 max, cognitive score, all these are features of that type. And all the other features somehow respond to this and become totally unstable. They are not able to recover. Their variance is infinite at about maximum lifespan, which is 120 years old. That means probably that the correct model of aging in humans is this. So on the horizontal axis is any variable that you associate with death, like C-reactive protein and something like that. Let this guy die. Die. Up. So you're born stable. You're born a naked mole rat. You experience stable dynamic fluctuations, stochastic fluctuations of your body weight and everything else. And then something else that is acting on those systems that is actually aging disturbs the regulatory potentials in linear way so that at some point fluctuations will tip you over and start killing. As Jan have explained today, actually, we kind of introduced this into the field, this idea of linear degradation of regulatory potential as the effect of something that is actually aging in humans, actually is translated to exponential acceleration of mortality. So, which means now, we need to find out what is actually driving. So I mean, in this picture, this linear cluster of features is true aging that is driving the degradation of regulatory interactions everywhere else. This is aging in humans. So that's what we need to target with drugs. And we need to understand what it means. So the question is, what is this? target with drugs and we need to understand what it means. So the question is, what is this? And for those who care, the property when your average is going up linearly and your variance is going up linearly, this is a signature, really the hallmark of a stochastic process. Most of the time, that means that this linear feature is just a combination of infinitely many unrelated, uncorrelated processes, and that's what biologists call damage, right? So this is damage. So how does it work? So in your body, you have many, many systems that can be in a healthy state and a pathological state. So for example, this is your DNA, and this is your mutation. This is your epimutation, or this is diabetes, and this is not. And these things are separated by very large barriers. I mean, that's why we live for a very long time. But what fluctuations do is that they activate people from healthy states to pathological states. And the longer we live, the more systems will go from healthy to pathological state. The total number of such transitions is just going up linearly as we age. The effect of each such transition is small, if it's not cancer, of course. But since the total number of such transitions in different cells is very, very large, together they produce a linear stress on all our regulatory potentials in such a way that these potentials become shallower and lower. And that's why the physiological variables that are essential for our functional shift, that's why we see aging, that's why my nose is going up, by the way, in size. And over time, fluctuations make systems tip over and leave the stability basin and start disintegration. We actually measured that. We looked at the data from wearable devices and we measured the fraction of people as a function of age who cannot return to the equilibrium within three months. And the total number of the relative, the ratio, the percentage of these people, the percentage of unstable people in the population is actually going up exponentially as we age and the exponent is exactly right. So essentially we measured that the number of people in this range goes up exponentially as it should be from the Gompers mortality rate, which is, to my taste, quite a demonstration that the overall picture is correct. So the problem with this, I mean, first it's nice, wow, now we know what is the true cause of aging in humans, uncorrelated damage. But the bad thing is that this damage correlates to the configurational entropy, which means that whenever you see that something correlates to configurational entropy, you start reading a statistical physics textbook. And the best statistical physics textbook starts from, can you read this? Ludwig Boltzmann, who spent much of his life studying statistical mechanics, died in 1906 by his own hand. Paul Ernfest, carrying on the program, died similarly in 1933. Now, ladies and gentlemen, it's our turn to study statistical mechanics. I mean, it's really that hard. So what does it, I mean, it's really that hard. So what does it, I mean, let me briefly tell you how this linear feature shows up. And for this, I will turn back to statistical mechanics. As you know, our cells generate energy. And they don't do it for fun. They do it in order to activate and deactivate multiple molecular pathways that are required for our survival. in order to activate and deactivate multiple molecular pathways that are required for our survival. So what these pathways are doing, they get activated when they need it and deactivated when they are not needed, and for that they use energy. Now, that bad statistical physics, dangerous statistical physics textbook is saying that there is no way in nature to translate energy to work in a totally reversible way, which means that some of this energy will be heat. And heat is the energy that you cannot use for activating your pathways anymore. Okay, we are evolutionally sophisticated animals, meaning that most of this heat will be actually dissipated, but some of them will stick into the system in the form of damage. And this damage will go up linearly as we age. So in signals like DNA methylation, that will look like this. So you have, for example, this is your DNA, and these are your different sites that are getting methylated or demethylated. So some of them will change methylation in a reversible way to activate and deactivate pathways. But this process will generate heat, and this heat will attack the rest of the molecules in your body in such a way that some of them will get damaged. And since this is kind of useless stochastic noise type of energy, it will damage your molecules, for example, DNA molecules in different cells, in a totally uncorrelated way. The total number of this damage will go up linearly with age. So that's why in any signal, you will see something that responds to stresses in a reversible way, this part. And then there will be another part that would be totally stochastic uncorrelated and on average linear. This is exactly what we see in humans, just to do it shortly. So this is the reversible part. Remember smoking part that is reversible. And this is the irreversible part. And that's exactly what we see in mice as well. So I mean, all animals are the same. There will be always a linear part that is the uncorrelated damage. And then there will be pathways that in mice are exponentially unstable, one of them, but I mean, they are reversible. Remember, I've shown you the example that you can actually revert, even the exponential pathway you can revert. How to prove that these guys are actually uncorrelated? It's quite easy. If you have single-cell methylation data, you can actually look at single-cell methylation samples from,ilation samples from different cells of aged animals. And for every site, you can ask yourself a question. Does this site know about the mutilation states of other sites? Meaning, is the mutilation of this site correlated to others? So for every site, you can find this measure of correlation. Low correlation means that this site changes its state without knowing what other sites are doing. And high mutual information means that this site changes its state with other sites. So obviously, these kind of pathway sites are moving together. They should have high mutual information. And these uncorrelated changes should have low mutual information. So we arranged all those sites on a line according to this mutual information. It then turned out that those guys, those sites that belong to these pathways, actually have high mutual information, so they are moving together. And the sites that are associated with this linear feature don't have high mutual information, meaning that, I mean, you really, from the data, you can show that the sites that contribute to this linear signature are actually microscopically uncorrelated, which means that this is just boring, the most dangerous kind of damage, because it's uncorrelated. So how anti-aging drugs actually affect these entropic and dynamic components of aging? This is not yet published, this is just a preprint that we have submitted this year. So look, for each animal, we can measure now this dynamic aging that we can change with drugs, and the entropic aging, which is most probably irreversible, and this is the true aging in human bodies. Take parabiosis. Parabiosis is this inhuman experiment where you connect young and old mice together, and young mice become older, and old mice become younger. So what happens to this dynamic biological age in mice? So this biological age is small in young animals. It's higher in old animals, as it should, right, because it's a biological age. And then if you connect old animals to young animals, this biological age, so young as a chronic, as they are detached, you can see that their biological age of the biological age is is reduced young as a chronic detached is reduced so this is for the for the true biological age and the difference stays by the way with time so you see that after detachment the biological age so it's reduced after the experiment and the difference two months after the experiment the difference Remember, you can do rejuvenation in mice, right? So, parabiosis reduces biological age, and two months after the treatment, the difference remains. Remember, mice are unstable. But if you do it, if you compute how the parabiosis affects entropy in these animals, the result is, like, no. So, young animals are still younger entropy wise than the older animals but parabiosis does nothing to the entropy so that's why i brought up this example with statistical mechanics so the common wisdom is that whenever you see entropy i mean what is entropy entropy is many microscopic states contributing to the same uh this many microscopic states contributing to the same microscopic states. Since our medicine cannot attack microscopic states, we can only affect states of many cells of many molecules with drugs, we are not able to change the microscopic states in any way, which means that most of our drugs will not work against entropic aging. And remember, entropic aging is the most important part of aging in humans, not in mice. So with this, I can show you one more example, which is not yet published. We looked, we're now looking at different animals. We want to find animals in which aging is more like human aging rather than mice aging. And one great example is dogs. So dogs are living longer than mice. Probably everything that lives longer than mice ages in a way which is the same as in humans. And in dogs, we also have a linear biological age, not exponential. You can fit exponent here, but the exponent would be negligible compared to what you need for mortality acceleration in dogs. And here we have now two different drugs tested in these animals, and none of them statistically significantly changes the rate of aging or even the biological age, the true biological age, the entropic aging in these animals. The drugs work in other pathways, but the effect of those drugs disappears once you stop the treatment, like smoking in humans. So basically my, I will be ramping, I will be finishing now. Basically my statement is this, that we started studying aging in mice, and mice aging is different from aging in humans. By optimizing drugs that extend life in mice, we actually optimize for drugs that produce reversible change in humans. So we're doing counter-smoking mimetic drugs, right? in mice, we actually optimize for drugs that produce reversible change in humans. So we're doing counter-smoking mimetic drugs, right? The problem with these drugs is that they affect in stable animals like me and dogs is totally reversible. Once you apply the drug, you have an improvement, but once you stop doing the drug, the improvement is no longer there, and they do not affect entropic aging whatsoever. So the conclusions are usually this, that we have different aging phenotypes. So in mice, you have this kind of inflammation-dominated phenotype that is also relevant in humans in the last 10 years of life. It's related to diseases. Most of longevity biotech companies are going there. They're testing their drugs in mice. They're getting great drugs that will probably help in diseases, they will probably extend human life by 10 years, and probably some of them will win, at least in combination, X price, but this will not change the rate at which your IQ is degraded, at which your VO2 max is degraded, so in a way, they are inviting the future that nobody wants, so that imagine that we will push off all our diseases, live to 120 years old, otherwise, I mean, I've shown you the data that we cannot do better than that, but your IQ would be essentially zero. And your VO2max would be essentially zero. So I think people understand what kind of trap the industry is leading them to, and that's why most of the people don't want life extension. And then there are good guys who are thinking that this is fun, but not for me, that there is another thing that is strong in humans, works all life, irreversible, because I respect physics textbooks, and I don't want to follow these two guys, those two guys. And if affected, it can produce multiple-fold life extension. So that's the only way to actually go. So basically, you can separate all longevity companies into four categories according to belief and ambition. So some believe that aging is reversible. Some believe that aging is not reversible. And some believe that you should do radical life extension. And some believe that you should do radical life extension, and some think that you should do incremental life extension. So interestingly enough, pharma is a life extension engine already, right? They don't believe that aging is reversible. I checked it. And they don't care about radical life extension. They want to cure one disease at a time. And they are very successful in that, by the way, already. Then there are a bunch of longevity biocompanies who believe that aging is reversible, but let's do small things like hallmarks of aging and everything else. Then there are elder slabs and other people who say, forget about small things, let's just do complete age reversal. And then there are people like Giro who don't believe that aging is reversible and still want to do radical life extension. So as you know, physics is the law, everything else is a recommendation. Now we have a great physicist to cite about that. And which will happen, I think, is that since aging is not reversible, these guys will either die or become pharma companies. I mean, there is really no other way around, which means that they will be able to do it