AI and Cancer: Unlocking the Immune System (with Immunai’s Luis Voloch)

AZEEM AZHAR: Hello and welcome to Exponential View with me, Azeem Azhar. The world is changing at an amazing pace and we’re entering the exponential age, a fundamental rewriting of our economic and social order that’s catalyzed by some remarkable radical technologies. Now on this podcast, I want to bring some clarity to the complexity of this change by helping you understand the trajectories of these technologies and their implications for business economics and our daily lives. In today’s discussion, I turn to the multi decade quest to cure cancer, and in particular, the crucial role in that played by our immune systems. During the past two years, living under the shadow of the COVID pandemic, popular awareness of our immune system has grown remarkably. More and sometimes less informed discussions about cytokine storms, T cells, B cells, neutrophils, and more have proliferated across social and traditional media. Meanwhile, scientific understand that the immune system has proved commercially valuable. Eight of the 11 bestselling therapies are immune centric, but that commercial potential is more than matched by the complexity of the problems and immunotherapies remain difficult, expensive, and slow to produce. Immunai is a startup that is hoping to change that by leveraging the exponential improvements in computing and AI on one hand and biological techniques like genetic sequencing on the other. Could this approach lead to a new class of immunotherapies that are both more effective and quicker to bring to market? My guest today, the co-founder and CT of Immunai, Luis Voloch, believes so. Louis, welcome to Exponential View.

LUIS VOLOCH: Thank you for having me, Azeem.

AZEEM AZHAR: Now we’ve heard a lot about immune systems, but I do think that we need to start with some basics. In my day, when I was a child, the immune system was pictures of white blood cells gobbling up a bacteria. And there was that famous film, The Fantastic Voyage, where people got shrunk to the size of cells and sort of fought off bacteria with many lasers. It’s obviously more complex than that. So, perhaps you can just give us a basic 101.

LUIS VOLOCH: Yeah, I’ll be happy to. So, first I’ll start with a disclaimer that I am not an immunologist by training or by any means. But I think the basic way to think about the immune system is as this army which is there to defend your body from all kinds of threats, and these can be bacteria, viruses or cancer as well. And this army has a variety of players. So, these are like B-cells, T-cells, NK cells, NKT cells, macrophages, and dendritic cells and there’s this huge A to Z of immunology. And each of them are building blocks of the immune system and they play slightly different roles depending on what they have to do. And I think one of the things that’s very hard for people is that there’s a lot of words and there’s all these different cell types and they sound exotic. But the immune system, essentially, is this army that has different divisions and each division plays a different role.

AZEEM AZHAR: I was really struck by the fact that it is a complex system in the sort of theoretical term of a complex system. It’s like a network of networks. There are different subsystems with different attributes. There’s the generalized system, which I think of in military terms as like the Marine Corps. They are pretty good at everything, but they’re not very specialists but you can get them there immediately. And then there’s the more specific systems which are armies and troops waiting in barracks that are very, very precise and can tackle particular diseases, but take a little bit longer to respond. And then within that, there are all sorts of other players in this complexity with very, very different roles. Do you think in terms of these different networks that interact in different ways or is there perhaps a different way to add some more understandable granularity to it?

LUIS VOLOCH: Think about it in very similar terms. So, you can think about the immune system as essentially a set of cells, as if you want to visualize it as like a tree. Essentially it first branches between the adaptive immune system, which is, as you said, are things that really evolve like your body and the innate immune system. And within each of those two departments, there are many different cell types and subtypes of those cell types and subtypes of those cell types and eventually states in which these cell types can be. And the big challenge in understanding the immune system is in understanding what is the role of each of these different sort of like many dozens or perhaps even hundreds of cell types do, depending on the threat that’s coming and more importantly, how are they failing to do that once they see on some kind of progressive disease actually arises.

AZEEM AZHAR: You describe it a little bit as a tree, as a sort of hierarchy, but often with hierarchies, the relationship is sort of linear, up and down, but is it like that? Or is it more in a sense, a kind of network where there are horizontal interactions that we need to be aware of?

LUIS VOLOCH: Because of evolution, the way that cell types eventually end up differentiating from each other and also as the bodies evolve, there is this hierarchical structure between the cells, but as you mentioned as well, a lot of what the immune system does when fighting disease, is acting as a network where you might have one cell in one part of this tree actually collaborating with a different cell from fairly different part of the tree, which it might, in turn, be sending signals to another cell in a fairly different part of the tree. So, there is definitely this network of cells collaborating to fight the disease.

AZEEM AZHAR: To what extent do you think our knowledge and understanding of the immune system has reached a sufficient level where we can now exploit that knowledge using advanced techniques versus us needing to still be on the voyage of discovery? I mean, I think sometimes about techniques that we might be able to use industrially and you are able to hit a point in the science where it’s sort of good enough that you can now build engineering on top of it. And there are other techniques where actually you are still rather imagining the science. And I think back, perhaps there is that things like nuclear fusion. What is your kind of reading about the sort of scientific knowledge and preparedness?

LUIS VOLOCH: So, one of the things that constantly shocks me is actually how much of the core immune system of understanding that we have around T-cells and things like that was just discovered in the last few decades, probably even in the last 20 years. And right now in biology, we’re really going through this revolution of more advanced and more high throughput methods, including single-cell genomics. And I think this will really change our understanding of biologists.

AZEEM AZHAR: You and your co-founder, Noam, got together and decided to set up this business, Immunai. Neither of you are immunologists. So, tell us what problem you set out to tackle.

LUIS VOLOCH: So, when Noam and I started the company and we’re both machine learning people at heart, our main goal was to improve the lives of patients. And in the beginning, we sort of quickly understood with the help of our scientifical founders, Ansu and Danny who come from cancer immunology, was that the main frontier in understanding why these cancer drugs weren’t working that well was immunology. So, taking a step back, cancer is a game, essentially, between the immune system and the cancer cells. And the characterization of the cancer genome, there had been a lot of progress there and there’s like lots of companies that profile the tumor. They would do DNA sequencing of the tumor and all kinds of things, but there hadn’t been yet a company that did the full characterization of the immune state of the patient and how it related to that tumor. So, Noam and I quickly, really understood the potential of that. And together with Ansu and Dan, our scientifical founders, we move into the next step, which was understanding what were the right technologies to be able to tackle that.

AZEEM AZHAR: So,  the idea of the immune state of a patient is what? It’s the mix of the cells across this complex system and the states that those cells and other molecules are in?

LUIS VOLOCH: That’s exactly right, and you can think about it in two steps. The first is almost as a pie chart, what is the breakdown of cells that you have both in the blood as well as in the tumor? So, for example, do you even have have T-cells in the tumor? Because if you don’t have that, it might be harder to fight a cancer because they’re some of the key players in fighting the cancer. And after you have this pie chart-like characterization, you want to go a step deeper into each of the categories and understanding which state those cells are. So,  for T-cells, are those T-cells still capable of fighting the cancer or are they what’s called exhausted? So,  I think being able to look across all different cancer types and find the similarities and differences of what’s making for good and bad responses was what we set out to do.

AZEEM AZHAR: Is it a fair summary, then, to say that your approach is one that is patient-centric? It starts by looking at the state of my immune system, whereas a traditional approach is based around the disease and the disease cause?

LUIS VOLOCH: Yeah, that’s right. We are relatively more like patient-centric in that sense.

AZEEM AZHAR: Right.  how does this connect to this sort of buzzword that has been flying around since I was at university 30 years ago: precision medicine?

LUIS VOLOCH: Taking a step back, the cancer has a few main pillars of treatment. The original one is surgery and then the two other ones that appeared over the 20th century are radiotherapy and chemotherapy. And in the last, mostly really in the last 10 years, we’ve had what’s called immunotherapies. So,  these are drugs that, instead of trying to fight the cancer directly, they actually give tools to your immune system to fight the cancer. And the beauty of that is that these tools are typically not cancer-type specific. They actually work for a wide variety of cancers, and they’ve really changed practice, particularly for advanced disease. And in precision medicine, when it relates to immunotherapies, the big question is, how do you understand the immune system of the person? Because you’re actually treating the immune system of the person. And once you understand that perhaps the patient is lacking a specific kind of cell types, that they have less than they need, or they have too much of another one or a specific one is in the wrong states, then you can actually understand which drugs to give to the patient to modulate the immune system of the patient into the right position. And furthermore, a lot of the way that people measure how cancer is being cured is just by looking at things like survival, which essentially means, what percent of patients survive after two years, for example, or has the tumor regressed. And what we can do is actually go a bit more upstream of that and be able to say, for these drugs that we’re hoping we’ll eventually make the patient survive, are they even creating the right immune response? Because if they are not, it’s hopeless from there because your army doesn’t have the guns it needs to succeed. So,  being able to look at this slightly upstream version of response, which is centered around the immune response of the patient, is very valuable to us.

AZEEM AZHAR: It’s a very powerful story, but how do you turn that into practicality? How do you start to advance these therapies? What is the information you need to collect? How do you collect it? How do you even make sense of it?

LUIS VOLOCH: In Immunai, we believe very heavily in looking at human data as like the main driver of discovery for us, as opposed to looking at just in vitro or animal tests. And we rely very heavily on collecting samples from patients that are being treated with drugs. So,  this includes blood samples, tumor biopsies, and So, on. And what we do is that, then we profile those blood samples or tumor biopsies with our technology and then connect that to whether the patient has actually responded or not to the drugs and then based on comparisons between responders and non-responders, we then try to analyze what’s leading to that. I’ll give you a simple example. Let’s say that for a specific kind of drug, about 60% of patients don’t end up responding. We can then characterize what makes those 60% different from the 40% that did respond to the drug and that gives us ideas for therapeutic interventions. So, essentially, new drugs that we can create to bridge that gap.

AZEEM AZHAR: So, at some point, what we have is, you’ve got a bunch of input data on a patient by patient basis. That input data will be stuff that’s easy to get, like their age and their height and their weight, and more complex data that you will have taken through some sort of assay. And then you also have whatever interventions they’ve had, whatever drugs they’ve had and what the outcomes of those will be over time. What is the data that you are actually grabbing?

LUIS VOLOCH: The main technology that we… With Health Pioneer is what’s called single cell sequencing. So, this is in contrast to traditional sequencing. So, typically, when people talk about sequencing, what you do is that you take a bunch of cells and you put them through a process, and the result of that process is that you have the average expression of DNA or RNA of all those cells together.

AZEEM AZHAR: Right.

LUIS VOLOCH: And with single cell sequencing, what you can do is that, you can put tens of thousands or millions of cells, and what you get is actually the expression of the specific cells. And this really allows us to tie it back to those many different kinds of cell types. When you’re able to characterize those, you can really understand the full heterogeneity of what’s there not just looking at averages.

AZEEM AZHAR: So, you’re going to have to go a layer deeper for me now, because I’m curious. When you say you are doing the single cell sequencing, and I was curious about this, is that a single class of cells?

LUIS VOLOCH: What we do is that, typically, we try to be as unbiased as we can. So, we’ll take, for example, set of a 100,000 white blood cells from a patient and we’ll put them all through the process in an unbiased way. And because the data that’s generated about each of those 100,000 cells is So, high dimensional, we can actually break those 100,000 cells into 50+ cell types and then really understand how those cell types are changing over the course of treatment and relating to response.

AZEEM AZHAR: So,  the kind of technical hypothesis is that you can, through this single cell sequencing, you can generate a lot of data for a given patient that allows you to characterize them and their response to particular drugs, and then if you have enough of that data, you should be able to produce some kind of useful predictive model that says, this is the best intervention for a patient characterized like this, and if not this, then this and So, on, right? That’s a sort of general idea, but then get to some more specifics, which is, how many? How much? Is it eight people? Is it fifty-seven people? Is it as many as a million people?

LUIS VOLOCH: One of the beauties of being able to work with single cell data is that, somehow the end for us, unlike with the end of traditional genomics, which is the number of people, the end for us is actually the number of cells. And this allows us to get results that we really trust sometimes with this few as thirty people in a cohort. Whereas, typically, when you think about things like the UK Biobank or these like large genomics projects, you need like many tens of thousands or even like hundreds of thousands.

AZEEM AZHAR: That’s quite distinctive, because often you hear this challenge, because each person is actually so unique and special in their circumstances that you really… To have any differentiated intervention, you have to get your sizes really, really small. And that’s often been a problem with how we thought of personalized medicine fifteen or twenty years ago, where you talk about these very large cohorts, right? And now you are saying, well actually we can start to do this with twenty or thirty people with a much greater degree of patient specificity?

LUIS VOLOCH: So,  the twenty or thirty people, cohorts, for us, they are the minimum size that allow us to be able to validate, at a high level, whether the hypothesis we had was correct, and then you need more people too, but typically on the order of hundreds and not on the order of tens of thousands.

AZEEM AZHAR: What is the standing database, the data network effect or the data pool that you build that ends up being able to power your predictions? Is that from patient data that you are gathering or is that coming in from third party sources that you acquire?

LUIS VOLOCH: It comes from both of them and taking a step back as well, one of the foundational hypothesis of the company was that we would like to look at several different diseases that have to do with the immune system. So,  and we’re focusing on cancer to begin with and looking at various cancer types. And only when you look across various diseases, you can actually see some similarities and differences that would be missed if you had only looked through one. And the motivation for that came from machine learning, in that there is this concept of what’s called the transfer learning, which are methods that allow you to take machine learning models trained on a set of tasks and apply them to different but related tasks. So, the canonical example of that is for computer vision, where you can train algorithms to differentiate between the chairs and desks. And you can reapply that to differentiate between other kinds of objects, because a lot of the basic foundational building blocks of the model, they’re actually shared across various images, such as being able to identify an edge and straight lines and textures and things like that. And we, essentially, wanted to take that to the immune system where there’s, again, all these different cell types and pathways, which are largely the same, but they’re being leveraged different ways across diseases. And we wanted to power discovery across all diseases by studying all of them at the same time.

AZEEM AZHAR: I’m curious about what you’ve learned in the… When you’ve applied this across many different cell types and disease expressions, what are the patterns in those classes that you have been able to identify?

LUIS VOLOCH: What we find is that, as expected, most of the cell types that appear while trying to fight lung cancer or melanoma, they’re actually similar, because it’s the same immune system that’s fighting them. But what we find is that when we pull all these cells together from all these different diseases, we’re able to identify more granular cell types. So,  for example, there are some specific T-cell or states of T-cells that might be prevalent on one disease, but they might be a bit too rare in another disease that you would be met.

AZEEM AZHAR: And what do you do with that new information then, that you’ve identified?

LUIS VOLOCH: The first thing that we do is to relate them to clinical response of those patients. So, we try to understand whether if this specific rare cell type might be related to response or non-response, for example, and based on that, this can give us therapeutic ideas for how to try to create more or less of those cell types in the patients.

AZEEM AZHAR: So, can we turn to how you want to apply this technology? Is this a platform that is about the discovery of treatments for patients who are currently facing a particular condition or particular cancer, or is this about building a platform that develops a pipeline of drugs that can be applied for future patients?

LUIS VOLOCH: It’s more of the latter. So, what we’re trying to do is to amass as much immune related data as we can and create this body of machine learning models that produce immunology insights and based on those immunology insights, it create different drugs.

AZEEM AZHAR: It’s obviously a really fascinating and super hot area, right? We’ve got a lot of companies are looking at this problem. And what they’re saying is, they’re saying that what made this discovery process hard… Well, what made it really difficult was, biological systems are really complex, that means they generate a lot of data, it’s been really expensive to get that data, didn’t understand the importance of transcriptomics or the epigenome. And then once you start to gather all that data, it was expensive to store it and it was expensive to process it, and it was expensive to find relationships between it. What you see is the exponential curves of these two different technologies, one is all of the ones that relate to sequencing driving down in price, and we’ve seen the price of sequencing the full human genome drop from a billion dollars to maybe a couple of $100 in twenty years, and, of course, the things that have happened in AI, both in terms of better algorithms and software optimizations, but also the declining cost of computational power. So, those things have coincided in this double exponential collapse the opportunity to tackle these biological and medical questions in this new computational way. But the thing that strikes me about the blockbuster drugs in this space, we talk about these sort of drugs that sell $10 billion, $15 billion, they were all produced using these kind of pre-exponential platforms in a way, right? So, there’s been a lot of promise, but there hasn’t been a lot of metaphorical beef delivered yet. Is that fair? And if it is fair, why is that?

LUIS VOLOCH: I think a large part of the reason is just timing. A lot of these drugs, they were approved already over 10 years ago, which means that their clinical trials started five or 10 years before that. And right now within cancer, there’s the largest ever number of clinical trials and a lot of the drugs in those clinical trials, they do come from these more computational heavy platforms. My prediction is that in the next 20 years, at least half of the top 10 pharma companies, they will… Probably they’re just being born now. And I think we’ll see a lot of those new drugs made then.

AZEEM AZHAR: So, it’s a matter of timing? Essentially, it takes years for something to get approved, technology to develop for teams to come together, to be founded, to be funded and to get moving. So, how do you measure the distinctive benefits of your platform to the traditional pharma approach?

LUIS VOLOCH: The main way that we’re measuring this internally is by looking at the efficacy of some of our early targets and drug prototypes and comparing them with the drugs that are actually in patients now. So, the results are quite promising. So, in in vitro results, we look at what we call immune phenotypes. So, essentially these are like, how much of the immune effect that you wanted to induce in the experiment, is actually induced by the drug.

AZEEM AZHAR: So, I’m curious about this from a business perspective. The driver here is, literally, you’re going to find better intervention. So, at least 50% better. The driver isn’t one of compressing the time it takes to find them or reducing the cost. It’s really, you’re going to find better treatments?

LUIS VOLOCH: In our case, yes.

AZEEM AZHAR: Why did you choose to go for better?

LUIS VOLOCH: We chose to go for better because it’s what we really thought would bring value. So, I think, as you mentioned, there’s lots of companies that, once you have some kind of drug prototype, they’re going to help you make the best possible drug within few months. And we, eventually, will probably partner with some of these companies, but for us, we thought that the main advantage we had was in just identifying targets better, making drugs that will address those in a really differentiated way, and eventually decreasing the risks that those drugs will eventually fail in the clinic because they are already mapped out in such detail in these early development phases. So, it reduces the risk. Because [inaudible 00:25:23] eventually failing in phase three.

AZEEM AZHAR: Are you building your own therapies based on your own understanding of clinical opportunities or are you still going to be a service provider to people who already have all of that clinical information? The main pharmas, for example?

LUIS VOLOCH: We are more of the former. So, we have our own therapeutic programs that we are developing internally. For some of them, we might want to take them all the way to the clinic ourselves and for some, we might want to partner with the pharma company to take them forward. And we’ve had quite a lot of interest in these early partnerships around our drugs.

AZEEM AZHAR: Are you agnostic, then, about whether you own the assets themselves or whether you co-develop them?

LUIS VOLOCH: We are very conscious of the value that we’re retaining of the asset and that’s the main thing we have in mind, and if we believe that the value of our part of the asset will be higher if we partner with somebody, then we’ll do that. The new trend in cancer immunotherapy is combining different drugs together. So, if, for example, we believe that the specific drug that we have will perform even better when it’s combined with a drug from another large pharma company, then it can be a win-win for both sides to create a joint program for putting both of them together.

AZEEM AZHAR: How do you know that that combination would work? Where is the data coming from to tell you that out of all these drugs in the market, this combination may be effective in this case?

LUIS VOLOCH: So, this also comes from our platform. So, a lot of the technology that we developed is around understanding when you try to target multiple pathways or multiple targets at the same time, if you’ve had like an actual combination.

AZEEM AZHAR: So, it’s one complicated thing to figure out what is the best therapeutic and what is the one that’s performing twice, as well as the other things that are available? But you obviously also have to go through the regulatory and approvals process, and before that, of course, there might be synthesis process that you have to figure out a manufacturing process. So, when you look at the rest of the pipeline that you have to go through to serve lots and lots of patients, how do you look at that? I mean, what are things that you think your approach allows you to improve in dramatic ways?

LUIS VOLOCH: So, the main focus of our technology, So, far, has been for target discovery and candidate prioritization. So, target discovery means, which genes or proteins are you trying to target? And candidate prioritization means, once you have a bunch of different drug candidates for that, how do you pick between them? And the other opportunities more downstream of that are, as I mentioned to you, we’ve also done quite a bit of work on like combination selection, but even later there can be things like biomarkers. So, these are things that are on our roadmap for working on later down the line, but these initial ones are the main focus So, far. So, again, target selection, candidate evaluation and selection, and how to combine them with different trucks.

AZEEM AZHAR: Now, when we first spoke at the top of this podcast, you mentioned that you weren’t an immunologist, you were in fact, a computer scientist. So,  what are the skills that computer science gives you that has allowed you to look at this challenge in a new way?

LUIS VOLOCH: The main asset that we have is that we look at everything through a truly unbiased and data driven way. I think this comes first almost at a cultural level for us, because a lot of the… Both me and my Co-Founder and our scientifical founders hadn’t previously put drugs in patients before. We actually didn’t come with a lot of biases around what should work and what shouldn’t work. And this is both a strength and a weakness, so to say. And this permeates throughout the whole company, and we’re trying to just make the right decision and evaluate things in as much of a data driven way as possible. And this leads into the more technical part. So, we never have this big scientist coming and saying, “We’ll take this target because I believe in this target.” We actually compare the target across various different metrics and just try to do as much of an unbiased evaluation as possible. And this is like quite a different mindset.

AZEEM AZHAR: I guess that the point being that when experimentation is expensive and you have to do your discovery by throwing it into a cell or into a mouse or something, you need to have some kind of hypothesis, some working theory to guide those expensive experiments. When you can run lots and lots of practice runs, millions of them, you can explore the combination and possibility space, essentially, at the cost of a transistor cycle. You can take a empirically driven approach to your computation, right? And you can divorce yourself from having to use theory as a guide to some extent, right? I suppose one question though, is, do you then try to work out what that underlying practical mechanism or pathway is? Or are you happy having a complete black box in why that system we’ve identified has effect, we’ve identified that there is a strong correlation that is seemingly statistically replicable, but we don’t yet understand that pathway, but it’s good enough for us to go on because the effect is large enough?

LUIS VOLOCH: I would say it’s somewhere in between where we are okay with doing a large amount of experiments without quite understanding the mechanism, but eventually before really believing in a drug, we’ll need to understand the mechanism and we’ll need to explain that mechanism to the FDA. So, I think that the way that we progress internally is by doing as many iterations of in silico and in vitro experiments as we can, and letting those experiments guide us to what’s actually working. But eventually, we do need to understand the mechanism quite well, but not necessarily at every step on the way. I think what one interesting analogy here is that again, like tying it back to this concept of cancer in the immune system playing a game, a lot of what we are trying to do is to get to understandings which would not have been possible just with immunological intuition. So, a few years ago when AlphaGo beat the world champion, this happened with this move that’s called Move 37, which was So, shocking to world experts because no world expert would actually have played that move. And I think, maybe, some of what we’re trying to do here is to explore biology in such an unbiased way that allows us to come up with maybe, eventually, what’s the equivalent of a Move 37 to cancer, which is something that at the end, we do need to understand, but maybe we wouldn’t have gotten there just with immunological intuition.

AZEEM AZHAR: I’m just curious about whether we need to understand it. I mean, is the need to understand the biological mechanism not just another medical dogma? As it happens, the year is 2022, we don’t know how general anesthetics work, we don’t really know how aspirin works, So, there are large parts of medicine that are in quotidian use, where we have no clue about what the mechanisms are. So, if it works and if you can prove it a trillion times in simulations, why do you need to go through this hoop?

LUIS VOLOCH: I think at a conceptual level, I’m with you, and I think on a more practical note, we should go through that hoop because it will help get these drugs into patients quicker. Right now, both the FDA as well as all oncologists and scientists, they feel the need to understand the mechanisms of these drugs and for us to be able to help patients the quickest possible way, the most sufficient thing is to work through that process and eventually empower those institutions and those doctors to better understand their patients. So, I think we feel trying to work with the system, bringing innovation will be the most efficient.

AZEEM AZHAR: That’s a really practical response and given the field you are in, I think it’s important to take that practicality, right? How do we maintain the forward momentum? But I’m kind of curious about this institutional gap. This deeply computationally strong, empirical approach that you have may move much, much faster than a set of much needed and well tested regulatory processes that were designed on a set of axioms, we are fundamentally gate limited by the number of possibilities you could explore. So, is there a discussion between companies taking your approaches? Let’s say, we’re going to need to… We should change that set of expectations of ways in which the clinicians and the oncologists think about this in order to make use of this new, super telescope that you and your peers have been producing.

LUIS VOLOCH: Yeah, you’re absolutely right there. And the way that we think about this is that, even if we believe that our drugs are much more likely to succeed in other drugs, because we’ve done So, many more of these simulations and other analysis, one path we could take or we could argue for is for these drugs to being approved just quicker. But we actually believe that this is, first, not going to happen, and second, for the whole industry, it’ll be better if we can look back 10 years from now and see that some subset of drugs that did go through this process that you described, actually did have higher success rates. And I think being able to prove that through the institutions would be better than trying to argue, beforehand, that they should change already.

AZEEM AZHAR: I want to come back to Move 37. So, what seems to be happening within Go is that, the power of AlphaGo has started to help human Go players develop new theories of the game. Is there an effect like that’s possible in your domain? And what might that look like?

LUIS VOLOCH: I would say that this has happened in different levels. In some sense, what we’re trying to build here is both the board and the game at the same time. The lens, through which we see the board, which is the single cell genomics, and then the strategy for playing the game, which is essentially like machine learning, like guiding the experiments. And the lens through which we see the board has allowed us to see that, actually, some of the pieces on the board, meaning the cell types for us, they are not quite the way that they were previously characterized in the literature, when you actually see them through single cell genomics and machine learning.

AZEEM AZHAR: The summary of that, actually, is that you’ve started to get signs of a new theoretical understanding of these mechanisms as a consequence of this new tool, which is single cell genomics?

LUIS VOLOCH: Mm-hmm (affirmative). Exactly.

AZEEM AZHAR: There is now, in 2022, a growing number of companies that are bringing together the skills of the biologists and the oncologist with the skills of the computer scientist and the machine learner. But it’s not, by any means, a established path. And it’s always struck me that there’s a lot that you have to build culturally, there is also a lot you have to build technically. So, I’m curious about the kind of challenges that you’ve seen when you of think about the full stack of what you need to build and what kind of culture you choose to build, sort of a armor culture, biotech culture, or a tech culture. So, you mentioned earlier in our conversation that, this field is this combination of two different exponential processes, the first being sequencing costs decreasing and more and more biological data being generated, and the second being compute and machine learning. And that’s also the way we think about this, but the issue is that for this product of two exponential process to work well, you need to have as efficient of a bottleneck as possible between them. And for us, this is a culture of multidisciplinariness in the company. So, for us, when we’re recruiting, we are very, very careful to bring people both from computer science, as well as from biology, that just think about the other side a lot. So, give you one anecdote is that, our VP of Immunology, he just like coding and python at night for fun. And he talks about data structures with our Head of Machine Learning.

LUIS VOLOCH: For us building that culture of multidisciplinariness is very important and it’s very, very hard. And we’ve had to make some tough calls around not hiring people that were exceptional in their field, but we didn’t think they would fit into that.

AZEEM AZHAR: Multidisciplinarity is like a multi-syllable word that’s actually quite hard to say, but I think it’s going to be even harder to build. So, just what are one or two of the key learnings and takeaways about how you establish that multidisciplinarity?

LUIS VOLOCH: The main lesson that we’ve had here is that organizational structure is very important for that. So, the way that most drug companies work, even the ones that do have strong computational teams, is that they’ll have a Bioinformatics Department and the Machine Learning Department and an Immunology Department and there are collaborations between them. For us, we actually have this flipped the other way. Whereas the main unit of work are multidisciplinary groups with a specific technical or business goal. So, this is what’s called in tech, the Spotify model. And it’s a kind of org structure that’s much more found in tech than in biotech and that we’ve brought here. And for like one of our groups, for example, it’s about 20 people, and it’s something like six or seven experts in single cell genomics, five software developers, I think five or six machine learning people and other immunologists working and they’re all working for a specific technical goal, but their day-to-day colleague are people of different disciplines, as opposed to just being people from the same discipline in their department.

AZEEM AZHAR: So, what often happens in these interdisciplinary teams in tech is that they’re led by a product manager and the product manager has been given a business goal or a user goal and the product manager knows just enough coding to be annoying to the developers and just enough about marketing to want to rewrite the marketing copy and So, on, but knits it together in a sort of sense of that centrality. Who is leading these twenty-person cross-functional teams in Immunai?

LUIS VOLOCH: So, first, this is evolving and there isn’t a gold standard, right? But for now we’ve played with two different arrangements. The first is of having co-heads of the groups. So, one person heavier on the computational side and one heavier on the biology side and for groups that are very technical with very long-term missions, we found that this works. And for the groups that have more dynamic missions, we actually have a product manager leading that as well, alongside the biology and computational heads.

AZEEM AZHAR: And what is the relationship between the computational side and that clinical and biological side? Because at some level, one could say, the computational people will be abstracted away and pushed further and further down the stack and they’ll look like a glorified Amazon Web Services, because what you’ll want to do is you’ll want to say to the clinicians, it’s all point and click now, right? And that could also be commercially desirable for you, right? Because you’re able to scale out your insights and infrastructure, but it also changes the nature of the balance between the company, because one part of the talent then only serves internal customers rather than serving the mission. Like in Top Gun, we remember Pete Maverick Mitchell, we don’t remember the guy who was fueling up his F14.

LUIS VOLOCH: We think about that very hard. And the scenario to describe that, I think, is one of our main failure modes for the future. So, when we are designing these multidisciplinary groups, the whole magic happens when you have the computational folks thinking through the design of the experiments together with the biologists and not just coming after the fact to analyze the data and be equivalent of the service provider there. And we believe that this is very important because as a company, we are not here to become a one or two drug company. We essentially want to create dozens or hundreds of drugs in the long run. And for us, every piece of data that we generate, which means every experiment that we make is actually a part of a larger data asset that we should be able to analyze and mine in its entirety together. And what this means is that when someone is designing an experiment, they need to think, not just about the specific biological question that they want to answer, but also about how that piece of data that generated will be analyzed together with the rest of the data that we have, later down the line. And this is a conceptual difference, again, in the sense that in most pharma companies there will be lots of experiments being run and then if someone wants to come later and do a math analysis or analyze many experiments together, they need to go through a lot of hoops to understand where is the data? How do I get it together? And a few weeks go by before you can do that. And for us, this all happens by design in the sense that, whenever an experiment is being planned, the metadata is already properly structured to be able to be mined together with everything else, and the experiments are designed for being compatible with the other experiments we’ve run.

AZEEM AZHAR: I suppose, in part, what I’m poking at is that, both of these skill sets, both of the types of teams that you are trying to hire are in industry viewed as Alpha Dogs, right? They are both really, really desirable. They are both captains of the space shuttle in a way. It just seems to me like, it would be like trying to have two number nines on a soccer team and kind of explain to one they’ve got to play in a slightly different place. I mean, does it ever feel like that?

LUIS VOLOCH: Recruiting is a big part of that in the sense that, we are trying to bring the number nines that have gone through the pain of understanding that they’re missing the other number nine? So, I’ll give you the classical example of that, are the biologists that have a vision of this massive amount of experiments that they would want to run, but they have no way to analyze the data on their own. And likewise there are brilliant computer scientists that love biology, but they just feel So, distant from the experiments because, maybe, it’s done by a different academic lab and they don’t have the kind of day-to-day interaction that allows them to help design the experiments. So, actually, I think what allows us to do that is that, both sides, through our interview process, they have gone through the pains of understanding that without the other side really being a partner, their outcomes are not going to be as good.

AZEEM AZHAR: If these approaches are successful with cancer in twenty years or thirty years time, could these approaches be useful for other diseases?

LUIS VOLOCH: Yeah, absolutely. We believe that, in many ways, cancer is where the immune systems are failing in the most egregious way, and that’s a great place to start from. But we believe that the same kind of transfer learning vision that we have within cancer, will also generalize So, autoimmunity or even some of these aging related diseases that have a lot of like immune connections. So, being able to look across various immune related conditions is definitely important.

AZEEM AZHAR: Louis, it’s an amazing journey that you are on. You must feel really privileged to be on it. If you had a message for people listening about what the potential is, what would that be?

LUIS VOLOCH: I think one of my main lessons is, also very much I shared with my co-founder Noam, is that if you are a computer scientist or a machine learning expert, think about biology and try to learn as much as you can because it’s some of the most interesting work that there is to be done with your skills, and we think it’s the one that will help people the most. We feel very fortunate, both, to be able to work on it and to be able to work with such amazing people as well.

AZEEM AZHAR: Well, I can agree with that. Thank you, Luis Voloch for your time today.

LUIS VOLOCH: Thanks, Azeem. Thank you for having me.

AZEEM AZHAR: Well, thanks for listening. If you enjoyed this podcast, do leave a review as they genuinely help others find it. And I strongly encourage you to turn to my recent discussion with Reshma Shetty, the co-founder and COO of Ginkgo Bioworks. Ginkgo is one of the pioneering biotech firms of the exponential age. They combine machine learning and genetically modified microbes to create novel industrial processes that will underpin our transition away from the fossil fueled economy. This podcast is a production of E to the Pi I Plus One Limited.