Scientist Stories: Tal Einav, Blending Biophysics & Machine Learning to Predict Antibody Response

So hello, my name is Tal Einav, and it is a pleasure to be here today to give my first talk as an assistant professor. I recently did a survey of the world, or rather I asked Google and ChatGPT, what exactly is biology? And they answered both in words and in pictures that biology was people doing experiments in order to understand how living systems work. It's, of course, a very narrow slice of a broad field. But if you slightly change the question to what is the future of biology, the nature of the response changes quite a bit. All of a sudden, you see people that are working on laptops or these futuristic tablets, number crunching. Sometimes it's not even people themselves. And there's this sense now that combining machine learning with biology will help push us into new frontiers. And the impetus for this change is just the amount of data that we are creating in our lives. To give you a sense of the scale, if we think about 1 byte of information as the fundamental unit, just telling us which letter of the alphabet are we talking about, then multiplying by a millionfold, you get to a megabyte of information. This is the amount in 2 books. Multiplying by another millionfold, you get to a terabyte of information, all the books in the UCSD library. And going another millionfold up gets you to an exabyte, which is now the scale of humanity. This is if we take all the words ever spoken by humans from the beginning of time until now, all the world, everybody, that is roughly an exabyte of information. It's an astonishing amount of data, but it's just the spoken word. If we think about over time how much content we've created total, then we are talking about hundreds of thousands of exabytes. In this year alone, we're predicted to make 120,000 exabytes of information. This is like a million million UCSD libraries. It's an astonishing amount, and it's what's fueling all of our progress and innovation. It's the information that's making our cell phones and laptops better each year. It's the progress and discoveries that are letting us create vaccines against viruses, or cures against diseases. So it is awe-inspiring, but it also calls out for us to step back and ask some questions about how we utilize these data and how we think about success. Because for much of its history, success in biology has been tied to how much data you generate. The more, the better. And this is still largely true today. And so in this context, we are, of course, greatly succeeding. But we also need to ask a parallel question of how can we use this information, learning from the past in order to inform the future. And this is where my lab lives. I want to take the 120,000 exabytes that we'll make this year and make them as informative as possible. To use the capabilities from Erika's lab and the discoveries from Shane's lab and combine that with data from dozens of other places and learn how we can treat diseases. So I'll give more, uh, detail in just a minute, but I want to emphasize that when you're generating this amount of data, there are many other questions about how you deal with all of it, such as where do you put it all, and what about reproducibility? These might not seem as provocative as the other questions, but they're just as important. It was only last year that the NIH developed 8 large-scale data centers in order to store all of these data. And one of those centers is right here at LJI, led by Bjorn Peters. And if you look at Bjorn's work, he'll put in a lot of effort to take these data and make them accessible and publicly available. And then he'll use that information in order to write papers that no one else has done before, using this unprecedented amount of data. It's not the traditional paradigm where you make this massive data set and then write one paper on it and go on and make the next dataset and the next one. It's instead work that is going to change generations because it enables labs like my own to really understand how all these hundreds of studies talk to one another and how we can look for all the hidden principles that were hiding behind the sea of data. So it's not data generation, it's knowledge generation. And I'm really, really thankful for everyone who's working on these types of questions because it is from their efforts that I believe that we're no longer data-limited in biology. And really, the success of these machine learning approaches will depend on where and how creatively we apply them. So what are the biggest questions that we can do with this monumental amount of data? People working in these fields are highly multidisciplinary because we didn't have this crazy quantity of data even 20 years ago. So the skills, the knowledge, the tools are all being developed right now. And we talk a lot about multidisciplinary work nowadays. What you hear about less often is that being multidisciplinary takes time and effort as you gain each of these different disciplines. So people who are working in these areas often have these quirky scientific journeys. Mine is certainly no exception. I started learning math pretty much as early as I can remember. The first toy I was given by my parents was this 4-function calculator that asked me addition, subtraction, multiplication, and division questions. And the nerdiness pretty much spiraled out of control from that point on. In fact, it was in summer math camp in 7th grade that I first learned about programming, and I fell in love with this way of thinking and speaking code. And I started to carry around these tomes of how do you code in Mathematica, how do you code in Java, and reading them whenever I could. I learned about physics for the first time in high school, and I fell in love with it. I ended up majoring in math and physics at Rice University, and it was really buoyed up by my programming skills. I cannot overemphasize how much these different skill sets synergize with one another. 100 years ago, if somebody told you that the force on an electron was A over x to the 10th plus B over x to the 5th, you'd have to visualize what that means in your head. Maybe take a few minutes to plot a curve. Nowadays you can plot it in a microsecond and have a slider that shows you what happens as you vary a or b. It's like standing on the shoulders of giants, and every one of the mathematicians and computer scientists that have been developing these languages really take every step that we take and magnify it a millionfold. So I had really loved every step of my academic journey, and I decided to go to grad school. And this is where the story gets a bit unusual, because most people who want to take a year off between undergrad and grad school do it right here at this moment. They go and work for a year or travel or have some life experience, and then they come back in and dive into grad school. I decided that I wanted to do the first year of grad school and then take a year off so I could tailor it to whichever lab I joined. And so my tentative plan was to join some theoretical physics lab and then go do a master's in math for a year. That's my idea of a break here. And, and so I came to Caltech and I talked to every professor I could find. And I met Rob Phillips, who was a physicist turned biologist. And Rob told me that physics has the best tools, but biology has the best problems. And I loved it. I loved his personality. I thought it was a match made in heaven. And I told Rob, awesome, I totally want to join your lab. I will see you in 1 year. Because I still want to go and take my year off, but now that I knew I was going to be in a cell biology lab, which is what Rob does, I figured I had the mathematical muscle that I needed. I would go and spend a year working on programming. So I did. I joined Wolfram Research in 2013. I spent the year as a software developer. And I'll mention that I had already been programming here for 10 years, but there are levels to every skill. And in this one year at Wolfram, I probably increased 10,000-fold my ability to program, even though I started highly proficient. Because just to give you a sense of what it's like to work at one of these companies, we had a listserv where you could ask any question. I'm trying to code this up, how would you do it? And whenever you're proficient, you can take one of these questions, you immediately come up with an answer. This paragraph of code will solve your problem. But there were people who had been working at this company for decades, and they would give these one-liners that are just— they bring tears to your eyes. They're so beautiful and almost magical in how they even work. There's no better way to get better at programming than to see these types of examples and to live it and breathe it for a while. For a year. So this was a great experience, and with all this knowledge in hand, I came back to start my PhD in biology in earnest. And this was hard. It was a huge transition. Math, computer science, and physics did not prepare me for the jargon and the information, the newness of biology. It was thrilling, it was terrifying. In my very first group meeting, I asked what GFP was, and the whole lab burst out laughing for about 10 seconds until Rob takes pity on me and says it's green fluorescent protein, which you'd learn in the first day of Biology 101. But I kept asking and I kept learning, and by the second year in the lab, pieces started to click together. By my third year, I felt at home working on biology. And in my fourth year, another Caltech professor, Pamela Bjorkman, came to give a talk about her synthetic antibodies that she was making against HIV. HIV is this very devious virus. It not only hides in our bodies, it also has a structure that's quite unusual that makes antibodies bind to it very poorly. So what Pamela was doing was synthesizing these new types of antibodies that could bind hundreds or thousands of times more strongly to the virus. And it was incredible work. I was stunned by her presentation, and at the very end she tacked on this very simple model that she admitted couldn't describe most of her data. And I stood up and I said, I think this model doesn't do justice to your really great work. I think we can do a better job. And we did. We started the collaboration, and it was a light bulb moment for me because this project pushed me on all of these different skills. Pamela's synthetic antibodies required some polymer physics that necessitated some very gnarly mathematical integrals that we then checked with hardcore numerics. And of course, I was learning about HIV biology the whole time. And by the end of this project, I was hooked. I felt like immunology was a field where my specific skill set can make a unique impact. And so for my postdoc, I decided to go all in and join an immunology lab. And I joined Jesse Bloom's lab where we study viruses like HIV, influenza, and SARS-CoV-2. And it was this time a gentler transition. It only took me 6 months to go from cell biology to immunology. But in those 6 months, all I did was read 3 papers a day and then talk to my labmates about them, asking questions, brainstorming ideas. It was really great. I could not have asked for a better set of mentors than Rob, Pamela, and Jesse, especially Jesse, who gave me the time and the freedom to grow and learn and fail and then pick myself up and do it all over again. It was exactly what a postdoc was meant to be, this protected time to get into a new field, and it really shaped the future vision for my lab. So let's jump to the present, and I'll show you what the results of this journey has been. Because as we've seen, our ability to defend against viruses is critical to human health and safety. And at this point, we have done tons of research on these different viruses, but I feel that we're still woefully underprepared to gain insight from those data. Groups will still too often publish this one enormous dataset about a virus, and then they'll write one paper on it superficially analyzing some aspect of it, and then move on as if they've exhausted all of the richness of these data. So many of my projects have looked at influenza, where decades of research have really shown us how individual antibodies bind to individual viruses. But when we get infected, we don't just produce one antibody, we produce hundreds. And this process is highly dynamic. As new viruses evolve, new variants will emerge, they'll reinfect us, and they'll elicit new antibodies. As old viruses go away, so do old antibodies, and the whole process is very random. So every person in this room will have a slightly different antibody response. And the central challenge for my lab will be how can we take all of our heterogeneous responses and create the best vaccines that will elicit the broadest, strongest, longest-lived responses to them? Think about each year when we get the influenza vaccine. Because the virus is evolving every year and we have to match this year's vaccine to our current virus. But even when we do a perfect match between the vaccine strain and the virus strain that circulates, influenza vaccine effectiveness is still less than 50%. And to understand that discrepancy, we have to look in people and understand our antibody response, because we understand the virus side, now we have to complement that with the human side. And this is a challenging question because there's this big gap between two key scales in biology, namely how well we understand single antibody-virus interactions, which at this point we really understand in exquisite detail. Where a single antibody binds on the virus, how tightly it binds, how well it inhibits that virus from infecting our cells. And then there's this growing interest in how all of our antibodies collectively protect us against the different viruses that we encounter in our lives. So the ability to bridge between these two scales will let us understand what happens when we get infected or vaccinated. And so it really lies at the heart of a lot of immunology research. My lab will bridge this question from both directions. We'll use a bottom-up biophysics approach to relate the behavior of individual antibodies with the full repertoire response. And we'll also complement this with a top-down approach where we predict how the antibody response evolves over time without necessarily needing to know its composition. So I'll start by describing this machine learning approach, and then I'll circle back to biophysics because I really think it's the synergy of the two that is going to crack this field wide open. So each of us has an antibody response, and we can characterize that response by getting a sample of your blood because that's where your antibodies are. Contained. And so we can measure how well do you neutralize many different variants of influenza. Neutralization means that your antibodies bind to and inactivate the virus so it can no longer infect our cells. So higher is always better. The viruses here that are very high, they're the ones that you neutralize very well. You won't get sick by them. But the ones that are very low are the ones that you're susceptible to infection from. And if we vaccinate you, we can remeasure your response and see what has been the shift or the boost upwards from vaccination? At this point, we have carried out this process so many times in the young and the elderly, across continents, in humans, mice, ferrets, and we still cannot answer the simple question of when do we expect that someone will respond well to the vaccine? Because sometimes when we carry out this process, we see almost no change between the gray and the green curves. So the vaccine didn't do very much, and we don't know what features of the response correspond to this null response. So I collaborate with people like Annette Fox, who's one of the heads of the World Health Organization, and she's been gathering these types of data for decades. She's very excited by these machine learning approaches because humans have tried to look at what parts of this response lead to a strong versus weak response. And we haven't figured it out yet. So maybe the machines can pick out the signal from amid the sea of data. But I think that we should be even more ambitious than asking this question. Even if someone responds to the vaccine, I think we should be able to ask, well, what happens had we given them a slightly different vaccine strain? Because there are dozens of different variants of influenza that circulate each year, and each of them could be our our vaccine strain. But currently vaccine design is a bit of a trial and error process. We don't exactly know what the green versus orange vaccines will do, and that's why we need models to show us what do we think will be the effects of having one vaccine versus another. The time is right to be asking these types of questions because more and more datasets are emerging that look at the pre and the post vaccination response. These 3 studies right here are massive datasets. They contain hundreds of different antibody responses dozens of different viruses, and multiple different time points. And if we can understand how the response evolves in one of these settings, can we predict what happens in these other contexts? So the big goals for my lab will be to create these universal algorithms for every vaccine that we have a dataset for, so that we can predict how any specific person's response would have behaved to the specific vaccine. We will then also address the question of time, because as we've seen from SARS-CoV-2, you don't want to make this transient response that comes right back down. You want to make something that lasts for years, if not our entire life. And these two methods come together to create a new form of personalized medicine where we can choose the best vaccine for you based on your specific antibody response. And so if we had designed two or three different vaccines each year, then we can choose for each person what is the optimal vaccine for them to get and what would be the benefit from having 2 versus 3 versus more vaccines each year. These methods will pave the way to rationally design the next generation of vaccines, and I'm really excited about these approaches because then we can really understand how every piece of the puzzle contributes to our overall protection. And finally, I'll mention that one of the edges that I bring to this problem is that as a physicist, I'm searching for general principles. So I won't be happy with some opaque box that can learn the rules for how our antibody response evolves. I really want to peer in and understand what's going on under the hood. And so my lab will do this in two key ways. We will use the growing list of machine learning interpretability tools in order to peer into these algorithms and pull out what features dictate a strong versus a weak response. And the second point brings me back to my two approaches slide. Because my toolkit extends beyond traditional machine learning. And an equally important thrust of my lab will be to look at the problem from the single antibody scale. So recall that when we measure your antibody response, we're measuring all of your antibodies working together. But we also have these large repositories where individual antibodies were characterized against many different variants of a virus. And so work by myself and others during my postdoc has shown that if you know how single antibodies behave, you can predict their combinations. So we can start to understand how our full repertoires could look like from all of these different combinations. I also pioneered efforts to go in the reverse direction, to take the behavior of all of our antibodies and pick out what is inside and how did those individual pieces behave. So if we, for example, measure your full response against many different variants, making these landscapes that I showed earlier, then we can pick out which of these individual antibodies do we need to add together in order to create those complex phenotypes. Which lets us go back to the same question as before: what happens when we get vaccinated? But now ask it from the perspective of individual antibodies. How do they change at every step of this process? I really like these kind of pictures. It's sometimes, you know, a picture can just change your mind on how to think about a problem because it looks like a neural network, but one that is biologically inspired. Whereas these hidden layers generally mean some amalgamation of information that you can't exactly understand, in biology they're only semi-hidden because we can go in and measure your antibody responses and see if the predictions match reality. So I like these kind of methods, especially because individual antibodies are the fundamental unit of our response. So these methods contain the fundamental elements of the biological system that we know should be there. And they let us ask very fundamental questions like, are our responses dominated by one antibody or by many? It's a key question because if we only have a single antibody targeting to one site on the virus, as soon as the virus mutates over there, those antibodies become worthless. But if we had multiple antibodies binding everywhere on the virus, even if one site mutates, the other antibodies still neutralize it. So we're much more robust against mutations in this context. How can we elicit these types of responses that target multiple different sites? And finally, I'm very happy with these models because they're a step towards mechanistic models, where we understand the system both at the fundamental single antibody scale and at the system scale. And for me, I really think that to, to look at this system and to harness and understand our adaptive immunity, we need to use these types of multi-scale approaches that let us analyze the system at these different scales. And true victory in my book doesn't come from just getting more data, but really from finding these actionable principles that we can use to design better vaccines. I'll end by mentioning that as a computational biologist, one of the joys is that you can take inspiration from a specific system, but your results end up applying much more broadly. I wrote papers on influenza and HIV, or even influenza and cancer, because at some level, when you focus on the antibody response, you can integrate out some of the details of the system. And so my lab will be experts in influenza research, but will also leverage that expertise to not just look at other viruses, but also look at other systems where antibodies are the key players, or where the same types of computational methods can help push the field forward. I believe in tackling big questions, and to really make a difference, I think that you must make your work accessible. And my group will strive to also be leaders in this very important I'll mention why I am here at LJI. Because I spent January and February this year interviewing very broadly. I flew from place to place, not even coming home for 2 straight months, interviewed at over 20 different places, had multiple offers in hand. And what I learned was that there are many places that do really, really great science, but there's also different things that flavors of science that people do, the questions they ask, whether they're willing to tackle big questions. And when I came here to LJI, I read the papers, I talked to these professors, and I realized that we were on the same wavelength and that I could contribute to their mission. These are not just rockstar scientists, but they are people who believe in the power of computational biology and that a theory-experiment dialogue will help push the field. Forward. So I am unbelievably excited to start my lab here at LJI. And of course, there are plenty of groups in San Diego and abroad that I'm eager to collaborate with. I think that science is best done in a community, and the right group of people can change the world. And with that, I'd like to thank you all for your time, and I'm happy to take any questions.