Scientist Stories: Shai Shen Orr, Human Immune Monitoring

Thank you to everyone around the world for joining us today for the Immunome Lab Meeting. My name is Kristin Aboud. I am the science editor at the Human Immunome Project, and I will be your moderator today. Coming of age is an expression that is used quite often these days when it comes to artificial intelligence and its applications. Advances in machine learning have entered our lives in numerous ways, whether we spend time on chatbots or not. For one thing, AI-backed systems are becoming more and more integrated into the practice of medicine, whether it is in interpreting test results, helping diagnose disease, or determining our response to treatment. The use of AI in medical research is also an incredibly fast-moving field. The New England Journal of Medicine recently launched an AI in Medicine series, with the warning that it may be a bit like taking a snapshot of the landscape from a bullet train. Today, we are going to hear about how monitoring of the human immune system using AI is one of the many ideas that is now coming of age. But before we get started with today's presentation, I wanted to announce that the application portal for the 2023 Michelson Prizes is now open. These $150,000 grants are awarded annually by the Human Immunome Project and the Michaelson Medical Research Foundation to support the work of early career investigators that are advancing research in the fields of human immunology, vaccine discovery, and immunotherapy for major global diseases. Reviewers are looking for research topics that are highly innovative and widely applicable. The deadline for submission for this year's awards is June 11th. So please click on the QC code or link listed on this slide to learn more about the prizes or to apply. Please also share with anyone who you think might be interested. Please be sure to also check out the latest issue of the Immunome Report, which features an interview with one of the 2022 Michelson Prize winners, Noam Auslander of the Wyss Institute, who is using AI to uncover new links between viruses, bacteria, and cancer. And now it is my great pleasure to introduce today's speaker, Dr. Shai Shen-Or, a professor of medicine at Technion, the Israel Institute of Technology, and a co-director of Tech.AI, the Technion Center for Artificial Intelligence. Dr. Shenor is developing machine learning approaches to study the drivers of immune variation, particularly in the context of aging, and seeks to further immune-based precision medicine. He is the head of tech.ai.biomed, as well as the Zimmern Institute for AI Solutions for Healthcare, and is the founder and chief scientist of Sci- CytoReason, a tech company that is transforming medicine and drug development from trial and error to a predictable practice by applying computational modeling of human diseases. During today's presentation, please send me your questions using the Q&A function in Zoom. I will ask our speaker a broad selection of your questions after the presentation., and then we will have time for discussion. With that, it is my pleasure to now turn it over to Dr. Shai Shen Or.

Thank you, Kristen. Uh, let me share my screen. All right, you can now see it. Yes, looks great. Great. So real pleasure, uh, to be speaking to this forum and, uh, real excitement with the everything that's been going on with the Human Immunome Project. And what I'll tell you about today is our own research in our lab, and which I think has high relevance as we think about moving into this 21st century where on one hand, the experimental tools are here to measure the immune system. On the other hand, the AI tools are maturing to the place where we can really be thinking on large scale and how this affects human health. Just as a disclaimer, as Kristen said, I'm the chief scientist of Cytoreason. As well. And maybe just to get going, I want to take a step back and think for one sec on what this immune system is all about. Basically, this is a system that evolved for two things. One is to maintain our homeostasis, and the other one is to protect ourselves from enemies from within and external, which ultimately comes back to bringing us back to homeostasis. So it's, if you think about it, this is a sensory system. It's a really complex sensory system that needs to pick up kind of anything that changes and respond to it. I think about it almost like a sixth sense. And the reason this is important to think about is not just from kind of a basic science perspective, but also if you think about this from a clinical perspective, on a very, very practical level, this is a system that kind of monitors what goes on internally, but also changes because of what everything that it sees. And so practically, we struggle with this question, a very simple question you can imagine posing to physicians, namely, how's my immune system doing right now? Just to give a sense of this, and I think we've experienced this recently with COVID if I were to ask my physician, how's my heart doing? So another essential system like the immune system. But there's a lot that we can assess about the, you know, your quality of cardiovascular health. There's a lot of ways that we can intervene both from lifestyle changes to surgical changes. With the immune system, despite it being so critical to medicine and disease, we don't really know the answer. We don't really understand this immune system variation that exists among us. And I'll show you some examples of this because it exists, because this is a complex system, because it responds to the environment. And we don't understand how we change over time. Now though, ultimately what this amounts to, this lack of understanding really amounts to inability to answer this, what sounds like a simple question. We don't really have good metrics of immune health. And to give you just as, we'll touch on both how this is complex and how we're our own research that of others is I think pushing the envelope on trying to change this, right? So just to kind of give you a sense, if I were to ask my physician this, on the left-hand side here, what you see is the very oldie but goodie white blood cell count. So this is your percent and absolute number of 5 different immune cells, white blood cells that basically constitute 100% of your, of the cells that you can see in blood of your immune system. This is a clinical assay. It's one of the, you know, top 3 most commonly used clinical tests done today. It's been around since 1959. So that's when we started measuring white blood counts on humans. And really in many ways, besides it being automated, it hasn't really changed. And this is really in stark contrast kind of difference to our understanding of immunology that change dramatically beyond those 5 different cell types that we can pick up in blood. So for example, if you were looking at this, you know, as a physician would look at this kind of parentheses here of what the norm is, and there's an arrow here that denotes what I have in my test today. So just, you know, say your percent lymphocytes, that can easily cover more than 200 different cell types. That exist in blood. Lymphocytes is an aggregate name for many of them. And you can see in this plot, this is an actual data plot we measured on a cohort. I'll mention this cohort again later where we have 30— we have 90 individuals of 3 different age groups, young, mid-age, and older adults. And you can just see the yellow lines describing the variation in their percent lymphocytes between individuals. You can see they're highly, highly variable between individuals. And also that there's a marked difference here with age. And zooming in here, which we can, not using the white blood count, but we have technologies to look at finer and finer subsets of lymphocytes. And you can see that this high variation, so here for instance is your percent CD8 T cells, a subset of lymphocytes. And you can see that this high variation exists within CD8 T cells as well, as well as subsets of CD8 T cells, you know, naive, central memory, effector memory, and so forth, all of which are highly variable. The point being is that this is a highly variable complex system. There's a lot of these different types of cells. There's a lot of, a huge amount of variation between them, between individuals. Every such cell has its separate distinct function. And we actually look very different from one another. So actually getting that answer to that, question of, hey doc, how's my immune system doing? Isn't really simple. And you can also see here how this is basically some kind of a personalized homeostasis. It's a personalized state that our immune system sits in. It's actually a cellular fingerprint. If I came and measured this in your blood and then I came in a month later, I'd be able to identify you without any problem. It's unique to you, right? Or almost unique to you, if I would say at that resolution. When I look at you at any given point in time. Of course, you can see from the changes with age that it does change. And so we ask ourselves, and this is true of every kind of immune system change, can we explain this variation? Particularly, can we explain this variation that goes with age? Because if we want to answer that kind of question that the patient poses the physician, we need to understand whether this is just noise or is there some particular type of pattern that explains this variation that allows us to predict what, you know, what you would look like in the future and what does it actually functionally mean. And to do this, one needs to realize the limitations that have existed in immunology for many years. Namely, we didn't have the technologies to measure the complexity of the immune system in humans. And that has changed over the past decade or so. And so what we started actually back in 2007 is a study in collaboration with Mark Davis. I was actually a postdoc with Mark at the time at Stanford and have been continuing this study ever since. We started a collaboration where we decided to measure healthy humans over time. And what I'm going to show you today is results from what is probably at this point the longest running study of the immune system, a high-resolution study of the immune system where we're looking at really the smallest building blocks of what the immune system looks like and trying to understand what is this variability we see between individuals. Can we actually find this pattern that will allow us to build quantitative metrics? And so this is work that we published a few years ago led by two students in my lab, Yael Talpert, an MD-PhD student, and Ishai Pickerman, a postdoc. And we basically took this cohort, we call it the Stanford Ellison Cohort, consisting of 135 young and old individuals. The young individuals here were 18 to 40 years old. The older were 60 and up. And we had all their clinical information. They were all, they never had been diagnosed with any autoimmune disease. They haven't had any kind of cancer for at least 10 years and so forth. We have a lot of information on them. We're starting off when we collected the first blood sample in 2007. These were healthy individuals. And we measured on them every possible coolest technology the, you know, available at the time. Every year we added more stuff. I'm going to tell you first the story from the perspective of the cells because that's where things really kind of made most sense. So we're measuring these people from 2007. Every year they come annually, we collect blood samples from them and so forth. And so after 9 years of data, we can ask ourselves, how is their immune system changing? And this is what we did. Now, what I'm showing you here is just one cell subset. This is a cell subset called naive CD8 T cells. You can see it here on the y-axis. It's frequency. And on the x-axis, you can see chronological age. Now, I'm basically skipping to tell you that we can find many, many, many differences between young and old individuals. So many, it's almost ridiculous to just enumerate them all. But what we care about is not as much the difference between young and old because that's almost trivial. Is there a pattern to how old people change, right? Are all the old people one group or can we make some sense of this high variation that we see? And so what you can see here is one very interesting analysis where in the black line you see if what happens if I average all the individuals across and just look to see what happens to naive CD8 T cells with age. And you can see that naive CD8 T cells drop with age. This is in fact nothing new. We weren't the first to publish this. There are hundreds of papers telling you that naive CD8 T cells as you age drop down. But what is new, what was new at the time when we published this, is those longitudinal slopes, colored slopes of every individual in the study. And what you can start seeing is that different individuals, the rate in which you change your naive CD8 T cells is actually quite different between individuals.. And in fact, sometimes even goes the opposite direction than what you would expect, right? So, there's a big rate change to this. There's a big amount of variation here between individuals. And one can ask, what does this correspond to? Now, this isn't the only cell that we measured. In fact, we measured over 72 different immune cells in people's blood. And so, we can ask this about each one of those different cells. And we saw tens of different cells in which this phenomena happened. Basically that the average wasn't really telling about the changes that are happening in individuals and that there was a huge amount of heterogeneity in that rate. Now, but the advantage of measuring so many immune cells in one individual is you can start thinking about them not as one cell type at a time, but as the whole high-dimensional profile of what an immune state is of an individual. Basically all your cells together. Now, this has only been possible in recent years because we put the technologies in that allow you to measure so many different immune cells simultaneously. Something that, you know, back 15 years looked like, you know, science fiction. So the data looks kind of like this. You have different individuals in different years. And for every such individual, for many different immune cells, you have their what is the frequency of how much you have of that cell? And if you have such data, you can start thinking about trying to imagine, could I actually tether this dataset together to try and move our individuals not from just describing them on a, you know, kind of one time point or 9 years of data, but rather try to describe the entire trajectory from young to old of how the immune system would change, leveraging the fact that we have a lot of information on every such individual about their immune cell composition and basically trying to stitch it together. The idea here is basically, so, you know, we would love to wait 80 years to look at how your immune system's changing, except we're too impatient to wait that time. And so instead we're going to stitch it together and introduce a term called pseudotime. Basically the idea of something that emulates what waiting that time would look like. And we could only do so because we measured the data in so many different dimensions, like so many different cell types. Basically, if you do this, you basically feed a computational, an AI algorithm, the input of how many immune cells, which immune cells were measured in every individual at every time point that you visit. And you don't give that algorithm anything else. You don't give them the, you don't give it their chronological age. You don't give it the visit that the person came in, the year, nothing like that. You just give the immune cell composition. You basically get this type of a figure where the axes here are basically a way— they're called dimensionality reduction. It's a way of, of taking something that actually happens in 72 different dimensions and putting it on a single slide that we can explain. So, you know, don't mind the axes as much as to see that there's some kind of a curve here, a linear shift that starts on the right-hand side and goes up here and then to the left. And we call this pseudotime, right? And when we color individuals by their chronological age, as you see here, right, what you see is that basically young individuals end up mostly on the right-hand side and older individuals shift on the left. But chronological age breaks down as you shift from right to left, right? The color no longer is organized in the same way. And in fact, even in young individuals, you can find the occasional person who has an older individual that has an immune system that looks like a young person. So potentially the fountain of youth, if you like. So this is kind of what we were able to do computationally. Now, you know, basically I didn't feed the algorithm anything but the immune cells. So, you know, what does it mean? Well, now we can study this shape, this pseudotime that starts on the right here and goes to the left and ask, is it meaningful in any way? So the first thing we noticed is that we look at such individuals. You can see the person here with the red arrow. This is a person who in year 1 sits here and then moves here and then moves there. And so he's basically shifting positively along this pseudo time axis. He's moving up with between his first visit, second visit, and so forth. Right. So now, is that a unique attribute? Well, no. We looked at all the old people in our study. Everybody, now you can see this in the pink distribution here, had a positive slope, right? You can see they were highly variable. So, but basically this means all the old people are moving from right to left over years of study, right? So everybody's, as they all advance on this pseudotime axis, their immune cells change over time. The only difference between people isn't if they advance or not, nobody's going back really. It's if are they, is what's the rate in which they're moving? Some people, are moving very quickly on this axis and some are moving very slow. Moreover, we realized, and this is what you see here on the right, that as you shift from pseudotime, from low pseudotime, we call it 0, so that's where the young people sit, to high pseudotime, which is 1, where the old, the oldest, the kind of the furthest data point sits. As you move from left to right on this axis I showed you here, your immune cells become your immune cell responses, the responses to stimulations of immune cells become more and more and more attenuated. They are no longer as responsive to challenges. The immune system seems to just be, you know, taking it easier. It's not managing to do the same response that younger individuals going. So basically what you can see here in a data-driven manner, just being able to measure your immune system in high dimension, we were able to identify this axis that appears to everybody's advancing on and it appears to be functional in the sense that As you shift on, your immune cells become more attenuated and it follows your chronological time, except it seems to be really decoupled from your chronological age. It's positively, noisily correlated with chronological age. You can see it here again with the color. So, different people are moving at different rates. Doesn't matter exactly what their age is. Now, the other thing to show you about this, and this is the same pseudotime, this is a heat map showing you the samples organized by the pseudotime here. On the x-axis, the samples organized this way. Every row is one cell type that goes into our algorithm. The column here is what the color denotes, how high or low is the cell. And you can see the young individuals have a certain profile. As you shift here along the pseudotime, you can see that the immune cells change very much. And in fact, here's the person I showed you before with the red arrow. In one year, he moves from this point this point, very, sorry, in 4 years from this point to this point. So very, very marked changes in what immune system looks like. And what we noticed in this is that if you look at any 2 rows here, there seemed to be an ordering between them. So you can see this over here where the, for instance, your CD28 negative CD8 T cells go up and only afterwards effector memory CD4 T cells go up. So it's not random noise by which our immune system varies. In fact, there seems to be a language between one cell affects the other and so forth. So, there's basically an ordering of the conservation of cell dynamics. There's a pathway by which our immune system ages and we can now basically pick it up. And we know that the different people just travel this pathway at different rates, right? Now, the bigger question to ask here is, is all of this clinically meaningful? Now, our cohort was very small. It was only 150 people and so forth. But we had a lot of measurements for these people. So for instance, we had gene expression data, your mRNA in blood that measured on the same people just as the cells I showed you. And so what we did is we learned a signature in mRNA and gene expression that correlates with that cell changes that I showed you along this immune aging trajectory. And in that signature, we now took to the Framingham Heart Study. The Framingham Heart Study is a very large study that's been going on since the 1950s. It has thousands of people in it. And it so happened that about 15 years ago, they measured on everybody gene expression just the same way we did in our own small cohort. And so for every person, we estimated what their position is on this trajectory that I showed you using this gene expression signature, we estimated the position. We call this position image. So for immune aging, so we can tell you what is the position in the trajectory based on either counting your cells or this gene expression signature. And when we did so in Framingham, taking all the clinical information that's been collected on Framingham since the 1950s, what we were able to show is that people that have a high immune age basically have a difference in survival probability to people who have a low image. And in fact, this is highly significant. So 10 to the minus 4 for survival. And it performs better than any kind of other biological clock that anybody has ever introduced, predominantly the epigenetic clocks that have been going on, that have been really popular. In blood, we can predict outcome by studying the immune system better than studying, for instance, methylation or kind of epigenetic changes. So, what did I tell you until now? I basically told you they have a clinically— we have an immune, a conserved trajectory in which individuals change at different rates. Their immune system becomes attenuated and it's clinically meaningful. We've been studying this in other contexts. Just to give you an example, in collaboration with Janet Lorden, University of Birmingham in the UK. We basically have looked to see that trauma patients have a different immune state, immune age than healthy individuals. Same goes for with cognitive impairment. So individuals that have normal cognition have a lower immune age than those with mild cognitive impairment or those with Alzheimer's. And more recently, Eric Feldman, a student in our lab, has been showing that in fact you can predict response to certain drugs at baseline by knowing what your immune age is. So this is in asthma. So to kind of summarize this from a take-home perspective, we basically showed that the immune system variation that exists in individuals isn't just random noise. It's actually a very conserved process, almost like a road, if you like, that we all travel from being young to being older adults, where the big difference isn't what the road, which road it is that we're traveling on, but rather the rate in which we travel on this. And we've been doing a lot of work to show that this is the case. And now we can take your blood sample and give you a score, which is the position on the road at any given point of time. We look very different from one another because we're a different place along the road. But if you wait long enough, one person is going to look very similar to another because we keep changing, and you get to the point where the previous person was, was there at a given, at a given point. So that's on the immune aging side. Except what we found out is that this is not just about immune aging. It's not about just being older adults. So we've taken this immune aging signature, our ability to position they get your IMID score. And we went to study Bangladeshi kids in Bangladesh, right? And what you can see here is a small study. This is still preliminary work, collaboration with Mark Davis and Lisa Weiger, where they published a dataset looking at children, 1 to 3-year-old children in Bangladesh. And what we did is we assessed how the, what is the immune age of these Bangladeshi kids alongside individuals from Palo Alto, And what you can see is that basically kids in Bangladesh have the immune system of individuals much older than the immune system age or immune age is much older than individuals than their age-matched children in Palo Alto. And interestingly enough, both children in Bangladesh and older adults in Palo Alto don't respond to vaccines as well, right? So, this gives us a sense of how we need to start thinking about modulating the immune system And with this, maybe I'll kind of take a kind of stop this kind of this phase of the presentation, just give you the philosophy of where we think about this. So what you can see here on the left-hand side is a very famous figure that was drawn originally by Conrad Waddington, a developmental biologist in the 1940s, that tried to explain how a single stem cell basically moves along a landscape to differentiate into different types of of cells. Could be a fiber, a skin cell, a muscle, and so forth. What we think is happening with immune aging is that your immune system starts at a certain composition, and as you age, it gets exposed to life, and that exposure affects your pathway. It affects how you travel, right? But the paths we travel are actually, you know, are quite conserved. The differences are mostly the rate changes, And ultimately, because this is your immune system, it ends up being that this relates drastically to kind of very strongly clinical phenotypes. Your immune state really affects what happens to you clinically and how you will respond to various immunotherapies such as vaccines or other immunotherapies. And that's kind of what we've been pursuing to date. So I'd like to Maybe shift gears here. Oh, one more thing. Just here's one slide summarizing this idea. Think about chronological age. And basically what we're telling you is, well, yes, you know you have a certain chronological age. It's very easy to ask you what it is. But if you want to be thinking about what happens to you biologically, you better study your immune age. And that is something that has a positive correlation with chronological age, but a very noisy one. We can now assess what your immune age and we can assess what, how this relates to how you were going to respond to different things. And so that's kind of what we've been pushing. But I, you know, as I speak about this, I want to also maybe shift to the next phase of my talk and say, well, look at all this immune stuff, all these different cells, all these kind of different measurements. How are we going to handle all of this, right? There's a lot of data here to figure things out. And so what I want to use the last piece of my talk to discuss is this idea of the data insight gap, right? Basically, since the mid-'90s, we've been exploding in data across every field, but particularly in biology. This has moved, you know, this keeps growing every day. And the problem is, how do you actually get insight from it? Right? Insight does not equate collecting the data. Insight right now in biology can easily be that I measured a dataset for, took me a few months to measure a dataset and I can sit years trying to understand what the data actually means. And so what we've been developing alongside our understanding of immune system variation are tools that are particularly geared towards immunology to get more insight from the data. Percent, think about like just getting the bang for the buck, the percent utility of insight needs to increase because every day looks worse here in terms of we're getting less and less out of every of our data because the data is growing exponentially and the insight growing linearly. This basically comes down to a technology we've been developing, which is the idea of combining many different algorithms together to introduce something we call the disease model. So basically the idea is very similar to what you see happen in the, phone or the navigation industry where we moved from having no data to having data collected but not being really highly utilized like your GPS, right? To actually having a model of the real world that uses this data in a continuous way and makes sure that, you know, that we're really getting the utility, the insight from the data. And this has transformed our lives. And in a similar way, we've been trying to do this in immunology where we shift from In the same period that phones changed, we're switching from having no data in a machine-readable format to having big data in immunology and molecular biology to trying to think about how do we take all this data together and put it in one framework that we call the disease model to actually change the rate and utilization of the data to a much higher degree. And so with this, I'll just show you a few examples here in the, Of this, we can think about the, what we call the data insight gap. And just a few examples of why this has been so difficult. So one side of this is the technology mismatch where we like, for instance, to think about cells, but the data, but most of the data that's coming out is not measuring cells, right? Or the fact that there's a lot of literature and we struggle to read it all. Or that we need to really integrate these together, or that we're now measuring, you know, individuals in a very, very rich way. They're very different from one another. But the way we've been developing our statistics isn't really looking at individuals. It's much about getting averages or the fact that every dataset that we learn from is standalone versus kind of building on everything that existed to date. We have no systematic way of integrating all this data together. So, with this in mind, what we've been developing are a variety of different algorithms to bridge the gap here. I'll show you the first one, this idea that cells are the atomic unit of the immune system and how we need to be thinking about them and how just thinking in cell space increases the resolution of what you're doing. We do this along this disease model that needs to bring in 3 arms, needs to bring in all the data that ever is measured on the disease I care about. And all the knowledge, all the literature about this, and do this in a repeated fashion so that we can learn and keep improving. So along those lines, we've been developing algorithms to really get the bang for the buck from the data. I'll show you one such example, or one or two such examples where in this case, for instance, we took inflammatory bowel disease patients who, for which there was data available in the public domain. These are 30 individuals who did not respond to anti-TNF, to infliximab, and 30 that do. As some of you may know, there are many, most drugs actually, less than 50% of individuals respond to them, right? And you can ask, what is it that, you know, can I predict before I give somebody a therapy whether they'll respond or not? There are financial implications to this. There are, ethical and clinical considerations of this. You want to get people the right drug and fast, right? These types of— this is 3 publicly available cohorts that have been around the public domain since 2009. If you analyze such data, you get these types of plot where on the x-axis you could see the differences, the full change between at baseline from the colon of IBD patients, inflammatory bowel disease patients, ulcerative colitis and Crohn's disease of people who were the full gene here between responders and non-responders to the drug. And this is the p-value. So, every gene here that I'm showing here is a gene measured in these individuals. And if the gene sits high, and that is a gene that we can pick up as being differentially expressed before you give people treatment, it seems like this gene is behaving differently between them. And there are hundreds, if not thousands of genes here in this list. And what we've been able to show is develop computational algorithms. So AI tools, if you like, that allow us to take this data and rather than look at it this way, start actually enumerating the cells in the data. So we measured mRNA, we measured gene expression, but we can build back what is the, cellular composition of those biopsies in these patients. And in this case, we're able to show that macrophages and plasma cells are at much higher abundance in non-responding individuals than responding. So, you can see that here. So, basically move from shifting about genes to thinking about cells, the fundamental unit of immunology. And when you do so, unlike anything that's been done before with these datasets, we're able to show that indeed, plasma cells and macrophages do in fact allow us at baseline to separate those individuals who respond to infliximab therapy from not. You can see here a validation of this in plasma cell staining. And this is done with an accuracy of over 82% on a held-out dataset from real-life patients in Israel. With the same idea in mind, I'm going to show you how it's important. So I kind of showed you how we need to be thinking about the data in a different way. But I can show you how we can think about the knowledge in a different way. This is work we did in 2019, and I think just shows the excitement and what we can even more now do with large language models. We basically built a natural language processing engine that's specifically tailored to thinking about the immune system. So it allows us to take a lot of text that's been published and identify the different cells there, how the cells are behaving, what are they doing, Which cytokines, which signaling molecules are they secreting? When those signaling molecules bind the cells, what do they do to the cells and so forth? We basically had a computer parse all this, put it in a machine-readable format. So now the computer can help us think about the entirety of the immune system across all disease and start asking where things match and mismatch and how it fits with our data and can we, bring learnings from one disease to another. Just to show you two examples, this is a graph on the left-hand side where every— the bar graph and every individual here, every column is one such signaling molecule, cytokines or chemokines. These are the signaling molecules which immune cells communicate. And you can see just the enormous amount of information that's in just this one single graph where The Y-axis tells you the number of different cell types that each one of those chemokines and cytokines actually interacts with, right? So the leftmost, for instance, is a cytokine called TNF. I mentioned it before when I said about the drug anti-TNF, that one signaling molecule can activate 72 different immune cells, right? And what we were able to do is just by extracting all the knowledge, bringing it together, we're able to suddenly take this kind of system-level view of what's happening in the immune system and realize that basically 23 cytokines and chemokines, which are only 17% of the literature of what we know about cytokines and chemokines, constitute 50% of our knowledge in the literature about what's happening in the immune system. Most of those cytokines and chemokines over here, we know very little about. But now we have a machine model, an AI model that allows us to start making predictions about what such interactions kind of fill in the blanks and try and predict novel interactions that have not been reported yet in the literature. That's exactly what we did. So for instance, we predicted based on all the literature and evidence that we've collected that monocytes should secrete IL-7, another particular chemokine, cytokine. And indeed, we did experimental— and we did experiments to show that this prediction holds despite the fact no one has ever reported this, neither in mouse or in human. We're able to now pick this up from the data, from the literature, and learn new and predict novel interactions. The same thing we can do for one cytokine across all diseases, we can do for one disease. So we can take every disease now and just bring in all the literature information and describe what is it that we know about the immune system in this disease? What are all the cells that play a role? What are all the cytokines that play a role? How do they interact with one another? You know, every edge here is clickable and I can bring in all the literature of relevance and I can do this for every disease that we know, right? And so I can do something now where I can take all the literature and start predicting new things. So for instance, what you see here on the left-hand side is a graph where every point In this graph is a single disease for which we have this network of what I just showed you in the previous slide of all the disease, the immune cells and the cytokines and how they interact. And we can put two diseases together if they seem to have a similar immune cell mechanism of action. The immune system works similarly in both of them. And we can start looking at both the similarities but also the differences. And so for instance, we were able to predict that the chemokine CCL8 should be involved in psoriasis based on all the other diseases that we can relate to psoriasis despite the fact that no one has ever reported CCL8 to be associated with the psoriasis. And again, you can see on the right-hand side two experiments that we pulled that allow us to show that CCL8 is indeed looking different in lesional psoriatic skin compared to non-lesional skin. So that takes the literature part into account. And just showing you the last piece of how one can think about the machine learning component here. Well, I'm going to show you a really nice application we did where we basically thought about any kind of human disease and how people often, and immunologists need mice to study this disease, to get at the biopsies, to get to the mechanism of action and so forth. But ultimately, once we work in mice, we need to get to understanding what's happening in humans. And the way this is done right now is not using AI. You take the mouse results, And you basically try and say, well, everything I saw in the mouse, more or less I expected to see in the human. And then it just ends up in the clinical trial failing when you realize that the pieces here did not in fact behave the same. And so what we developed is an AI model we call FIT for Found in Translation. And FIT allows us to basically takes in the prior of all human diseases gene expression data versus healthy control and all mouse models of those diseases, pairs them together. So you can say this is a mouse model of this human disease and this is a mouse model of this human disease, learns the relationship between mice and men so that if you give me a new mouse dataset, I no longer use it standalone, but actually I now leverage all the great experiments that have done before to help me fill in the gap. What am I expecting to see in humans, right? We're using the machine, the computer to actually bring that AI into the translation process from my mouse results. And long story made short, this allows us to increase the translation, the kind of knowing what is going to happen in human by up to 50% in certain diseases. You can see that on the left here where you can see what we call the similarity between any two mice, two human diseases. And that you can see here, that's what 20% accuracy. You can see what the mouse— when I try to move between mice and men without our AI model in blue here, and red is what happens when you, when you fill in, when you use the AI model to tell you what's going to happen in human, you basically boost your accuracy. And on the right-hand side is another experimental validation experiment where the model predicted something should happen in human. We, you know, there was no human data to show that this is the case, and we did the experimental validation to to show this holds. So, I'll end here with really two points. The first is I showed you this idea of immune aging as an image, as a quantitative metric to quantify baseline immune state. We all go through the immune system aging just at really different rates. It's a very conserved process. It matters both from a molecular and clinical side. And we've been developing the AI tools to get the bang for the buck. Where do I think all this ends? I think this is where we're heading to. We're heading to a future where each individual knows his or her immune health. No longer the CBC since 1959, but really measuring the right immune cells to know what your kind of where you sit along this kind of disease progression and health axis, understanding the interactions between different immune cells and having tailored norms that fit your particular your particular condition. And just thank you to the lab and the funding agencies.

Happy to take questions. Thank you so much, Shai. That's incredible. So much work and so interesting. I think you've stunned everyone because with so many things to think about. But I actually have a couple of questions for you. I'm wondering, how different does this look in people with autoimmune diseases?

In terms of the immune aging?

Yes.

Everybody's been asking me this lately and I don't know yet. Okay. But I believe— so we have running bets here, right? So we basically think— so we know in immune-mediated diseases such as inflammatory bowel disease, the immune aging is predictive. We're able to see that it allows us to predict disease severity. And then when you respond or not respond to a drug, you can modulate your immune age. For at least for up to 3 months, which is what we've tested with autoimmunity. You know, my, my own personal bet is that you're going to end up with a higher immune age. But we have yet to measure this. I'm hoping in the next quarter we'll know. We have a student doing this in the lab today.

Okay. That's very interesting. And is it, is it like almost an exhaustion of the immune system? And does that help explain, like when you look at the Bangladeshi children, because of their exposure, presumably to far more pathogens, right? Is it almost an exhaustion of the immune system?

So I think the word— it's a wonderful question. I think the word exhaustion is, you know, is overused a bit in immunology. It means several things, right? So it's not exhaustion how we think about T-cell exhaustion and so forth. What I think is happening is actually similar to what happens to the brain when it learns, right? You know, when you— if you— you know, there's horrible experiments people do with blinded cats, right? And they didn't This is years ago.

I'm not going to tell my daughter about those.

Yeah. So, so similarly, think about this. If you take the immune— the immune system needs to play a balancing act. The balancing act is between two things. How strongly does it need to respond to an acute response? Right. It got an infection. You want to slam it, right? So you want to do it with the full force. On the other hand, you get chronic exposure. It could be chronic exposure in Bangladesh. It could be you know, longer exposures to life as it happens in developing and developed countries, this process yields a lot of chronic things that you end up stimulating repeatedly. If you're really, really, you know, if you're gung-ho and doing these, the best, strongest response you can every time, you know, over with that chronic exposure, you're going to start hurting yourself, right? So the immune system plays this balancing act of needing to attenuate.. And in Bangladeshi kids, this happens early. And in older adults in Palo Alto, this happens late. And ultimately that optimization game carries a penalty. And the penalty is that you end up not responding as well as we wanted to acute responses to infections, to COVID, to vaccines and so forth. And so we're going to need to learn how to modulate it.

Right, right, right, right, right. Yeah, we just— and we just got a question actually. Can your older immune age be artificially sculpted back to a younger age with the right cytokine cocktails?

Uh, it's a wonderful question. Uh, I personally think, uh, yes, but it's going to take us quite a while to— we're studying this, but it's going to take us a while to actually understand what it is, especially I think the question isn't— so we already know we can modulate it with a drug, but the question is, is it going to stick over time? And the point is, you know, I can change your immune system, but I'm not changing your body. And that body is in interaction with it. So, you know, we somehow need to modulate them both. So we think we can do short-term changes with diet, with, you know, kind of exercise and so forth. Longer term is going to take us a while to understand how to move it. I also Maybe important to say, I'm not sure in every condition an old immune system is a bad thing, right? Think about we're studying in the lab, for instance, bone marrow transplant. So, in bone marrow transplant, you can think about the donor giving you their immune system and then you have that issue of graft versus host, right? Where the immune system attacks you, right? Maybe getting an older donation from that perspective would be worthwhile, right? So, there's a lot to play and think about here.

Yeah, sure. And we also had a question on how different immune aging might look in patients with cancer following immunotherapy.

Wonderful question. We are looking at this now. So I don't know the answer. We certainly before therapy, we've been hesitant to look from a perspective, you know, we think there's too much havoc going on. And so we're not sure we're going to pick up the signal. But post-therapy, our thoughts are that this would be a good way of trying to understand when you stop immune checkpoint inhibitors and so forth, as well as potentially we're just now starting a study to look to see, can this actually be prognostic for identifying early diagnosis of cancer? Right. We shall—

time will tell. Right. And do we know that the immune aging was one of the factors in the higher death rate with COVID-19 among the elderly?

Is that something you have? So we have seen a relationship with— so we don't know that per se. What we have seen, this is still unpublished results, is that we can basically predict COVID vaccine response based on immune aging. And in fact, you know, and that individuals who have higher severity also seem to have a higher immune age. But I, you know, I don't know this to the level of kind of survival and death and so Right, right, right, right, right.

Well, it's very interesting too, because, right, there are vaccines that have been found to work quite well in the elderly with certain adjuvants, right?

So it's, it's an interesting— That's exactly how we think. We think that you basically, it's the right combination of, you know, if you think about vaccines as targeting the specificity of the immune system, it's the adjuvant that allows you to modulate the immune system so that you will respond. And we think that immune, understanding this pathway, this conserved pathway, will allow us to actually do much more intelligent matching of adjuvants to vaccines.

Yes. Intelligence and vaccines. I love it.

Dare to dream.

No, but thank you so much. That was really— it's such interesting work and, you know, so important as we all age. So thank you so much. I really appreciate your time and want to also thank the attendees for participating in attending today's webinar and for submitting such great questions for our speaker. I wanted to let everyone know that our next Immunome Lab meeting will be held on Thursday, May 11th. So please mark your calendars now. Joining us that day will be Dr. Michael Snyder of Stanford University, who will be speaking on the topic of managing health using deep data and remote monitoring. You don't need to register to join, but please visit the website below to to obtain the Zoom link and for more information about both our past and upcoming webinars in the Immunome Lab Meeting series. And while you're there, please sign up for the Immunome Report, which is another platform for sharing the research and voices of scientists working at the frontiers of human immunology and AI. And finally, please do visit our website and follow us on social media where we will post of today's webinar. Thank you again for participating, and we look forward to seeing you again at the Immunome Lab Meeting. Thank you all so much. Thank you, Kristin.