Lessons from Ron Alfa, Co-Founder and CEO of Noetik, on Bringing AI-Native Precision Oncology to Every Cancer Patient

Speaker A: Hello everyone, I'm Rachel Glaser, one of the co hosts of the PEAR Healthcare Playbook podcast. On, um, this podcast we sit down with founders building category defining biotech and healthcare companies from idea to scale to learn about their experiences. Today we're joined by Ron Alpha, co founder and CEO of Noetic. Noetic is building an AI native platform for cancer biology, generating large scale human biological data to train models that both predict patient responses to therapy and uncover new drug targets and treatment opportunities. Ron, thanks so much for joining us. For those who don't know you yet, could you please tell us a bit about your background and what inspired you to become a founder?

Speaker B: Yeah, thanks Rachel. Great to be here with you. Let's see what inspired me to become a founder. Well, actually I never wanted to be a founder at the very beginning. I mean that wasn't my initial goal. My initial goal was to go to medical school and uh, find treatments for diseases. That was my initial inspiration. So uh, and really that came out of a frustration with you know, the progress in medicine. Uh, you know, like many people, uh, have had people my life that suffered from cancers and other diseases and you know, as you sort of interact with the medical system, you oftentimes get very frustrated and you're like, why can't we do anything here? So you know, that was my initial motivation. You know, really I felt that if I can get into, you know, translational research, then I can really make an impact for patients and you know, help to discover new therapies. Then you know, flash forward maybe a decade. I was in the MD PhD program at Stanford, um, and had the opportunity actually to meet Chris Gibson at uh, who's co founder of Recursion. And Chris and I worked together um, at the Ignite program at the GSB Graduate School of Business. And that was really kind of the first experience I had, uh, you know, in, in something that was like a startup environment. Light, you know, we were, you spent the summer working on kind of ideas for Recursion. Um, and that was really fun. It was, uh, you know, I love kind of working as part of a team and like working on like a hard problem and you know, it was very inspiring. So uh, so Chris left and went and started the company and then I went back to medical school and finished my clinics. Uh, and then you know, back in 2016 I called Chris up and I said, hey, would love to join Recursion. And I, you know, moved to Utah and joined the company. So that, that was, you know, that's a little bit of My origin story into startups and into tech bio.

Speaker A: Yeah, thanks so much for that origin story. And we definitely want to dig into it a little bit more. So maybe to start, uh, I'd love to hear about what first drew you to the intersection of computation and biology. And when did you start to feel like AI might fundamentally change how we understand biology?

Speaker B: Yeah, I mean, so when I joined recursion, it was 2016. So I would say we were at the very beginning of this wave and very early days in terms of maybe not necessarily early, uh, computation as a whole, but really early days of computation impacting biology. Um, and really what we could start to see even three years into the founding of Recursion. So Recursion was founded in 2013, but what we can really start to begin to see is the scale at which we can start to produce, uh, biological insights. And even then we were not even near the scale that we are now. Uh, but you know, really the, the fundamental question that we were trying to answer at the very beginning of the company was whether we can, um, essentially in an unsupervised way, characterize the biology of hundreds of different disease genes, um, you know, without looking at specific, you know, markers or biological pathways. But, you know, can we use computer vision to really begin to characterize genes across, you know, hundreds of different areas of biology and then begin to relate those to each other? So that was our sort of, you know, core founding vision is like, okay, you know, can we, can we look at, you know, I wouldn't say genome wide initially, but, you know, early on, you know, large scale, you know, functional genetics in an unbiased way. And even at that early time, you can start to see that, yes, like, this works. So we can use computational tools, um, to broadly characterize cells that have received perturbations, um, and, you know, do so in an unbiased way. Um, and that would also allow you to do, you know, ah, therapeutics discovery, types of screens in an unbiased way, focusing on specific pathways. And historically, the way we do biology is we narrow down into a very reductionist view into one or two things. But all of a sudden we can step back and we look at many different pathways using the same methods. And so it turns out that becomes incredibly powerful because that data set basically builds upon itself. So I'd say back in really the first days, uh, of Recursion and you were starting to see these threads that what we were building was going to be scalable and was going to start to unlock biology across, you know, a Large swath of essentially human biology.

Speaker A: And so maybe could you talk a little bit more about your learnings at recursion, how you think, how you started to think about building companies in this space based on your time there and also what convinced you that a new company, Noetic, needed to exist?

Speaker B: Yeah, so those are big questions. So let's see, I have a lot of learnings. Uh, I think one of the things there's sort of phenotypes of folks that um, now multiple of us that have left Recursion and started companies so you can sort of track the phenotypes if you will for ex recursion folks. Um, and one common phenotype is we've all deeply believed from the very beginning that um, in order to build uh, advanced computational capabilities in biology you really need to generate your own data. Um, and certainly there are some exceptions to that rule where there have been like large scale data sets in the public domain. Obviously PDB and AlphaFold is maybe the best example of uh, uh, capabilities that have leveraged 50 years of data accumulation, um, in a very structured and curated way. But generally speaking, um, across biology we just don't have the right data sets and you just don't have the scale of data to brute force these problems. So I would say one important learning that, you know, a lot of people uh, that have started companies out of recursion, um, and you know, really principle to the core to recursion was that you have to generate your own data, um, and you need to build the data really fit for purpose for what you're doing and that will unlock um, a lot of the capabilities. And so when I founded Noetic, you know, having spent many years at recursion thinking about how to build these data sets and we were in those years learning a lot, so testing out different assumptions, learning what works, what doesn't work, learning about batch effects, learning about confounds in the data. When we found the Noetic, six years after I joined Recursion, we could make some assumptions going into the data generation process based on a lot of those learnings. Um, and then I would say maybe the other key aspect, um, that spending six years at a company at the intersection of technology and biology you learn is that it's actually not easy to integrate a lot of these teams and capabilities and operationally think about how to build a company in this space. So all these companies exist in a regime where you have very different types of people that need to work together. Um, you have a whole spectrum of folks that are technical in different ways, from, like biologists to, you know, sometimes chemists and engineers and, you know, computational biologists and ML engineers, ML researchers. So you really have like, so many different phenotypes of people that often are not working closely together, um, that you need to really think carefully about how you like, construct the teams and how to bring people together. And, you know, also just, you know, make sure that, you know, folks are very like, mindful that like, everyone is bringing something really important to the table.

Speaker A: I definitely want to ask more later about how you think about recruiting talent, especially across diverse areas of biology and computation. Uh, maybe for now let's dig in a little bit more into Noetic's thesis and what you're really doing there. So you describe Noetic as aiming to build foundations, models of biology. Could you please talk a little bit about what you mean by that and how you're approaching this differently from other players?

Speaker B: Yeah, so a core thesis of the company. And this is maybe not obvious to everyone in this space, this tends to be very obvious to people that are, you know, deep in pharma clinical development. But this is not obvious to people even in the, in the tech biospace is that a lot of drugs fail in the clinic for efficacy. And oftentimes that's not because we picked the wrong drug target or, you know, that's not because they were not safe or because we didn't pick a good molecule. Oftentimes molecules fail because we just don't know which patients are going to benefit from those drugs. And especially in oncology, this is, you know, arguably one of the, one of the most important problems in oncology specifically, um, if not drug discovery as a whole, um, and so imagine, for example, you know, very obvious example. So, you know, probably everyone's familiar with Keytruda and immune checkpoint inhibitors, um, you know, some of the most promising, uh, therapeutics ever developed for oncology. And you know, these drugs only work in some small fraction of patients across many cancers. So, you know, for example, in lung cancer, Keytruda only works in about 12% of patients. If you were to enroll a trial with all of lung cancer and, you know, the, the drug was impactful in only 12% of those patients, that would be a failed trial. You, you know, that would not be successful. So the way that we've discovered, you know, we've been able to, to identify these drugs and moving forward is by identifying, you know, this subsets of patients, those 12% where, where these drugs are effective. And you tend to see this, you know, as we move beyond sort of the regime of chemotherapy. Most drugs, you know, in cancer and, you know, other diseases, most diseases are not homogenous. Most drugs will work in some fraction of patients, but where the biology is aligned with the biology of that therapeutic. And it turns out that, you know, I fundamentally believe that we are not terrible at, like, preclinical science. You know, we are not bad at running animal models. We're not bad at, uh, doing in vitro models. We're not bad at, you know, scientific work early on. But you find that there's this very, like, large disconnect between the work that we do before you get to the clinic and projecting who's going to be the right patient, you know, once you get to the. To the clinical trials. And so that creates a disconnect where, you know, many drugs fail because we just don't know how to enroll or design these trials to have the right patients. Um, and this is a huge cost for drug discovery. This is, you know, clinical development is the most expensive part of drug development. And, you know, drugs that fail in clinical development. So 90% of drugs fail, 95% in some cases. Um, you impact really the whole economics of the industry, and yet no one's worked on this problem. So, uh, you know, this is our contrarian thesis at Noetic, which is, you know, if you want to solve this quote, unquote, translation problem, and translation usually refers to, uh, you know, identifying which patients in the clinic, ah, a drug will work in, and then you really need to train models on the right data. And that data needs to be human data. So, you know, we're not going to bridge translation from, you know, more cell line experiments or more animal model experiments or, you know, better, better organoids or composite cell lines. You know, we've tried to do that for decades and decades and decades. We really need to start from, like, the source of the biology, which is human patients, um, and leverage innovations in, uh, the technology. So AI and, you know, innovations in AI architectures, um, to be able to draw causative biology from these hard to, um, you know, hard to intuit patient samples.

Speaker A: Since you brought up AI and combining that with patient data, uh, I wanted to ask about something you said before, which is that LLMs won't solve biology. Um, but could you please elaborate on why not? And what is the missing ingredient?

Speaker B: Yeah, I mean, there's a lot of hype and, you know, I'm famously on axe, uh, trying to dispel some of that hype. I think there's a lot of hype from the frontier labs around. Okay, you know, the next thing we're going to do is solve cancer, but yet no one is working on cancer. You know, uh, people are, are, are not, you know, these labs are, you know, maybe, maybe recently people are starting to do some work in biology over the last maybe few months. But generally speaking, you know, folks are not working on cancer. So fundamentally, you know, very simple, my very simple premise is, if so where models are good, um, is where they have the underlying data. So it's no surprise that, you know, models are very good at coding. There is mountains and mountains of training data for writing code. You know, we've any software engineer that's learned to code has spent many hours of their life, uh, on the Internet trying to figure out how to uh, you know, build the next version of their thing. So there's tons of that training data. I would argue that the training data for solving or understanding cancer biology or any disease biology just doesn't exist in a large enough form, um, to be able to train models, um, to really get to novel biology. And I know there's some idea that, well, once we get to AGI, then the models can reason against data and they can do scientific experiments and they can train themselves, so on and so forth. So you know, I'm optimistic about that thesis. But here we are today, um, there are people suffering from cancer, we have this technology. I, um, would argue we should build the data sets fit for purpose for solving this problem. We should train the models to solve this problem. Um, and I think we will be able to leverage, uh, reasoning models, natural, uh, language models in conjunction with those types of capabilities. So for example, know, I think, you know, there's a lot of power to the, the protein models, uh, and you know, I think those are incredibly valuable for helping us identify molecules and design molecules. But I think the max leverage comes if we can combine those with, you know, reasoning models, AI scientists, agentic frameworks to do that type of work. And I think we're going to see the same thing with uh, biology foundation models. I think we're going to need to build biology foundation models on the right biological substrate of data. Um, and then those models in their embeddings are going to understand fundamental biology and they're going to help us make new discoveries that are just going to be too difficult to make by bridging the broad disparate data that's been in the public domain.

Speaker A: So you bring up this idea of using data that's generated fit for purpose for machine learning and this data probably has different design parameters than what you might say is traditional scientific data. So could you talk a little bit more about what that looks like in practice and what kind of specific parameters this data should have?

Speaker B: Yeah, I guess fundamentally I think of this and maybe I'll not uh, even say fit for purpose for machine learning. Like, let's just say if we want to build a foundation model of X, um, whatever that is, we need to think carefully about what is the data that we think is going to solve this problem. Um, maybe that's obvious, but I think that's not often what is done. I think oftentimes, uh, folks address this problem from the standpoint of some data set that already exists or some m way. We've been doing things for a different approach. So for example, um, in biology, we've been generating scaled computational biology data sets for a long time. Um, and it might feel like the natural next step is for these data sets to be machine learning datasets. Um, but that's not necessarily, it doesn't necessarily follow that, um, the large data set that you generate for some computational biology experiment is necessarily a good machine learning dataset, or it doesn't follow that that dataset is necessarily going to tackle the problem that you're solving. Um, and I won't get into any necessary specifics around that. But what I'd say is I think it's important to think about, well, what is the problem you're solving and what data do we think is going to be able to enable us to solve this problem? And you're not going to be right out of the gate. You're going to have to learn and figure it out and iterate. Um, but at least start from the very beginning and try to understand, okay, well, what are the data that we think are going to hold the biological insights for this? And then also to understand the sort of constraints and confounds of the data set that you're generating and try to control for some of those things. So for example, when we founded the company, um, we started really from first principle and said, okay, if you just wanted to solve this problem, problem of building a biology foundation model that understands human biology, what kind of data would you want? Um, and you want to be able to train a model, the model needs to see biology, relevant information, that relevant information needs to be translatable to humans. What type of data needs to be scalable? Um, and so as you start thinking through this problem, um, you look at all the types of data that you could begin to generate and you start Thinking through, well, okay, what do I want? Well, you probably, if we're dealing with humans, you probably want something that um, allows you to access human biology in the clinic, you know, where patients, you know, see doctors, where patients get diagnosed with disease. Um, and so for us we said, okay, well one, one disease, one layer of that data could be the pathology stains that, you know, every cancer, um, gets diagnosed by. Um, so we can, we can generate that data, but we know that those, those images are not, are really not that rich in terms of biological contrast. It's just a couple stains, you know, they, they are, you know, rich in contrast. But you don't have molecular labels. And if you really want to understand biology, you probably want to get to some level of, you know, molecular pathway genetics. You know, you want some molecular labels. So we can go one step down and say, well, um, what if we started looking at cells? So instead of tissues, but if we're looking at, also we're looking at cells. And so we said, okay, well let's generate a protein data set. Um, protein data set still doesn't get you this scale of molecules. Um, and so then we said, well, how about we look at transcriptomes? Um, and because we want this data set to be, we want to build transformers, it turns out transformers are very well trained with images. Let's make all these datasets images. Um, images are also very cheap to generate, generally speaking, um, they're very scalable. So we can generate the scale of data that we want, uh, to train these large, uh, foundation models. Um, and we can also get replicates across many different patients. So you start to look at all these features across your dataset and you begin to optimize this, uh, multi parameter problem of what is the right data. Um, and then you sort of hit go and start generating data. And so we did that. And you know, actually we couldn't train a model for, you know, about two years. Uh, you know, we're not going to train a model on 21 patients. Um, so we, we basically set up a lab, we opened the lab, we bought machines, we started generating data, you know, sourcing patient samples. And after about two years we, you know, had a few hundred patients and we could start training our first models. And then we actually learned something. We actually learned, you know, initially we thought ok, uh, the first models we trained really largely depended on the protein layer. Um, and we trained those models and they worked, but they were really hard to interpret. We didn't have that many protein states. Um, and so it was really hard to understand what they had learned. So then we sort of shifted the effort and started training the models, principally focusing on the transcriptomics layers and thinking about things in the context of cells. And, and then all of a sudden those models were very easy to interpret and we could understand what was going on. Um, and so that changed our perspective and we learned something. And so I'm kind of giving you this example of like, well, you start from first principles and you'll go down one path. Um, and then you have to iterate, you have to learn, okay, well, is this working? Is it not working? Um, and then you could end up down a totally different path.

Speaker A: Maybe just to follow up a little bit on this sort of multimodal massive dataset that you're generating. I think, you know, two things that are really interesting to our listeners is like, you're collecting all of this data that involves like, single cell transcriptomes, spatial transcriptomics, imaging, histology. And you talk a lot about how important, for instance, spatial biology is to the noetic thesis. So what does combining all of this multimodal data really unlock that wasn't available previously?

Speaker B: Yeah. So we've started to see a lot of benefits from spatial context. And, you know, it's, it's not hard to intuit that that would be important. So, you know, think of, think of a biological experiment where you just have cells isolated and let's just pretend you have, you know, two or three types of cells, um, and they're isolated and you know, you, you can profile them and you, you can look at their transcriptome, um, you know, one cell type at a time, there's a lot of information there. Um, and, you know, if you're isolating them from an experiment or even from a tissue, you know, probably their, you know, the RNA signature of those cells will vary from, you know, depending on where what, where they were taken from. And you know, that information is embedded in the cellular biology and, and you know, it's capturable, but all of a sudden you've lost the context. You actually don't know. You know, if I have a bunch of immune cells, I don't know if those immune cells came from inside a tumor or if they, or if they came from a normal tissue. All I know at the point of experiment is I have a pool of immune cells and I have a pool of tumor cells, and I have a pool of normal cells. And so there's a whole layer of context that is lost. Now you can imagine the spatial version of this experiment where if I just have two cells Right next to each other. And I'm profiling, let's say they're transcriptome and let's say I'm just looking at a thousand genes. So increasingly we're looking at you know, 19,000 genes. But let's say I have two cells and I'm looking at a thousand genes. There's information in those genes, the patterns. So the model can learn from the sets of genes that are represented by each of those cells. But there's an information layer that you're, that's sort of not obvious to you, but it is the relationship of the, every gene from cell A to cell B, um, let's say the distance between those, the signals. Um, and you know you can imagine a cell very far away that has another thousand genes. And again you have information in the relationship of your, you know, gene one to, you know, to cell type C, if you will. So there are spatial, there rich spatial information contained across, you know, those, those layers of biology, um, that actually are very rich uh, information for the model to learn from. Um, and we found that you know, that type of information is actually really important for scaling these types of models. So one of the things we showed recently was um, if we want to um, for example train larger and larger models with larger parameter sizes, um, that actually these models um, um, do better with larger context windows. Um, so the context window being how much of the tumor, uh, you know, the model is getting exposure to.

Speaker A: So let's pivot a little bit to Noetic's business model and maybe talk a little bit about the JSK deal which I think a lot of listeners are going to be excited to hear about. So when you think about Noetic's long term business model, is it as an AI licensing and data platform play, building an internal pipeline or some sort of combination?

Speaker B: Um, I would say our long term vision is to solve oncology and initially, maybe initially to solve oncology but ultimately to build foundation models of biology. I think the question a lot of companies in this space are asking is as we start to build these models, um, how do we build a company around them at the same time? How are they useful? Um, and so I think it's hard to appreciate today that that was also the same question that all the frontier AI labs were asking, let's say four years ago. Uh, it wasn't necessarily obvious what the business model was going to be for some of the LLMs as they were init being developed. But it turns out as you start building these models and start deploying them, then these business Models start to emerge. And so broadly I would say our focus is on how do we train models that really understand cancer biology, um, and then ultimately understand human biology, um, that can really be powerful tools in the service of getting better drugs to patients. Today we're at a point in the company, so we're three and a half years in, um, we've generated a quite large, um, proprietary multimodal data set, um, around thousands of patients. We've trained our first models. Um, and we're finding that these models are incredibly powerful. So you can use them for, um, you can use them to run simulations. These are world models. You can run simulations which you can use for, you know, therapeutics discovery applications. So finding new targets, you could find them, use them for using, uh, for doing biomarker discovery. You can use them for solving clinical trial, you know, a, uh, response prediction if you have patient data. So these models work. They're really powerful for a host of drug discovery applications. And so naturally the first business model has manifested is, um, companies that are really excited about the space of oncology and really devoted to making breakthrough drugs, um, have begun to reach out to us and say, can we partner with you, um, to leverage your models in our portfolio? Um, and so what was unique about the GSK deal in that respect was that um, it really was the first um, you know, AI bio, ah, foundation model licensing deal. So really we've been thinking about these types of deals for a long time. You know, I've spent many years uh, at recursion, even before founding Noetic. And we've really always been thinking about, well, is there a business model where one can license, you know, a model, as it were? Um, and I think historically it's been, you know, harder to transact these. I think, you know, generally speaking we haven't, you know, models are just getting to the point where they're very broadly useful. Um, but, you know, what's, you know, what's I think enabled this uh, type of uh, partnership, you know, in our case for this model is that, you know, really one, we're training these models on the types of data that every pharma is working with. So if you talk to any company that's got uh, an oncology pipeline that's working in clinical development, they're all generating this type of data. Everyone has translational teams that are bringing in patient samples that are profiling these patient samples with protein stains, with H and E, with spatial transcriptomics. So these are the types of data that, you know, everyone recognizes our core to understanding Human biology and are important. Now all of a sudden we've trained, you know, a model, a foundation model on these types of data that allows you to unlock the value of these data sets. And it turns out that, you know, these data sets have been really difficult to work with at scale. Um, there just haven't been very many great examples, um, of being able to leverage these translational data sets to really make a huge impact on clinical development. And I don't want to, um, ah, understate, you know, the great work that people on these teams are doing in translational departments. Certainly, like, there's a lot of work in biomarker development and um, that's been a core part of drug discovery for a long time. But you know, arguably most people would not say that today. Um, you know, biomarker development is, you know, sort of the, the biggest game changer in a drug program. Um, and so what we're imagining is if these models can really begin to understand biology in a fundamental way, then all of a sudden the power of these approaches gets to the point where it is, you know, the deciding factor in whether a drug is successful or not, which is, you know, incredibly powerful and has the potential to make a lot of impact across many different methods.

Speaker A: Yeah, thank you. How much do you think of the JSK deal? Sort of reflects the unique value of noetics data and the models versus also a broader shift in pharma's willingness to pay for AI instead infrastructure, which as you mentioned, is something that's also happening very much in the field.

Speaker B: Yeah, uh, that's a really good point. I think both of, I think both of those things are true. I think we are seeing, you know, a shift towards, um, pharma wanting to pay for AI infrastructure. You know, we are, you know, I think everyone's recognizing the importance of, of AI in the space. Um, and if you, if you go look at any one pharma company, you know, they are working on dozens, hundreds of programs across many different indications. Um, and I do think that we're getting to a point where companies want to be able to leverage these capabilities not just in the context of one focused research collaboration, but really across their whole portfolio. Um, at the same time you have very different capabilities from molecular design models to AI scientists to bio foundation models, um, and everyone's not going to build all of these capabilities. Certainly some of the pharma companies that are, um, at the frontier of AI may build one or two of these capabilities. Um, but generally speaking, building everything in house that has never been Pharma strategy. You can even think of their core business which is making molecules. And historically pharma has, you know, outsourced a lot of that work, even, even licensing the molecule, uh, you know, themselves, once they get into clinic. So you know, I think building things in house is not sort of necessarily core to every company strategy. So there are, they are going to turn to companies that are, you know, experts in building these capabilities, uh, to be able to license these capabilities. You uh, know, one of the challenges is there hasn't really been, you know, really a deal structure for these types of, of transactions. And so from a, if you're on the business development team, you know, it's, it's not obvious how you structure these and there are complexities to it. You need to think about, well, what are, you know, what are we licensing? You know, what is the thing, what is the model, um, how do you license the model, um, how do you protect it? And so I think there are, you know, some complexities that uh, you know, will, will make these, you know, go slow at the very beginning. But as you know you start to see more of these transactions, they will become a lot more commonplace. Um, and I think, I think the floodgates will open, you know, maybe this year or next.

Speaker A: Yeah. And you know, when you think about noetic smote in the context of this, where do you think it lives? Is it like the model weight the data? Is it the ability to source human patient data which is often hard to access because of regulatory and consent barriers? What do you think?

Speaker B: Yeah, I mean, uh, it's interesting. It's always hard to think about moats in these types of spaces. Certainly we are not in the space where we're all training on the same data set. So we have a very large data moat. It's taken years, millions of dollars to build this data set. We've sourced all these patient tissues ourselves. We have complete freedom to operate around everything from the biobank to the data. Um, we have all the patient samples here in the lab behind me so we can go generate additional. Um, so there's a tremendous moat in both the samples and the data themselves. Um, but the other thing that's unique about our data set is you just don't have a public domain data set out there that looks like this. Um, and so I won't um, say that folks couldn't build similar types of models to us, uh, if they had access to these data. But we've really had to figure out things that are quite unique to building these multimodal models. So what are the loss functions? What are the masking policies for the, uh, MAE models? How do we tokenize these different data modalities? And we've iterated across multiple different model architectures at this point over the course of years. Um, and so if all of a sudden there was a very large data set out in the public domain, certainly folks would start to figure out how to build these types of models. But currently there just isn't a data set of this type of, um. And so I would say we have um, quite a nice head start around how to construct these types of models. And um, we're seeing that they are incredibly powerful and working.

Speaker A: Yeah. And to follow up on that, a classic challenge for companies like Noetic is that you need this expensive proprietary data set to build a model that you can license and that's worth paying for. But then you can't really fully prove the value of that data until you have it and actually make useful predictions. So how did you decide what data set to build before you had proof that it would work and what were like early signals that it was working before external validation?

Speaker B: Yeah, I mean, that's uh, a, this is a uncomfortable question because, um,

Speaker A: M.

Speaker B: You don't know that it's going to work unless someone's already done it. And this is always the regime you're in when you're doing something new. So if you're the nth version of a company that's following, you know, like the, the, the 10 companies that came before you in building a certain type of model capability, um, on a data set that exists, then you, you might have a good intuition about how you can do things differently, you know, slightly differently, and then make an incremental advance and, and you know, have, you have a good intuition around, you know, at least that the core principle of your business is going to work. Um, if you go out and you're saying, well, we're going to build a foundation model of biology on a data set that no one's ever built before, um, then you're, you're, you're flat into the regime I'm of, I'm starting a new company with an idea. Um, and we don't know if this is going to work. And you know that this is how, you know, people, hundreds of people start companies all the time. So I think to some extent you, you just don't need to accept that, uh, you know, this is how you, how you do new things. You know, this is how you do new things in science. You design experiments you test them now at the very beginning. I think it's important, as we were talking about before, it is important to think about well, what is the end result you're trying to achieve. And it's not sort of random chance that we've designed the data set in this way. We really have thought, we thought carefully for many months. We talked to a lot of folks that have done this before. We talked to folks in the regime ah, of digital pathology. I've spent many years in the space of high con imaging and cell based models. So there's a lot of sort of information that has, you know, been integrated into the decision point of what is, you know, this, what does this data set look like? So it's, it's certainly not random. Um, but you know, I would say there's always a bit of risk in that. You know, you just have to start generating the data and see um, and see how it's working along the lines. Um, now what are the indicators? So we started, we intentionally started generating data in lung cancer. Um, and people often ask why lung cancer? Well, turns out lung cancer is a space where we have some handholds. Um, first of all we know we can get a lot of data in lung cancer. We know a lot of people have studied lung cancer um, for many years. But you have some biological handholds. We know some patient populations that you can hope that the models can learn about, um, from self supervised training. Um, so for example I mentioned Keytruda and the immune checkpoint inhibitors. We know that there's a large fraction of patients, large 12% of patients that respond to Keytruda in the lung cancer population. So a question is, can we find those, do the models learn those patients, um, you know, without labels, um, just from self supervised training? Um, uh, we also know that there are different subtypes of lung cancer that respond to other therapeutics. Do they learn those, uh, subtypes. And so if you're as you, as you start from the beginning, if you're in a regime where at least there's some known biology where you can test, then as you start, you know, training your first models, you can begin to ask like are they learning the things that we expect them to learn? Um, and that, and then that can help guide you. Um, and then maybe the last point here is I think you just have to be kind of optimistic in a way. There's sort of this trade off between and it's really difficult. So if something's working like this, you have to recognize that it's like actually 30% working and not 0% working. Because if it's 30% working then you have to keep pushing in that direction. But if it's 0% working, you need to switch directions. So it's really important to be able to be a bit optimistic and realistic about whether your first versions of something look like they're directionally correct, um, so that you can make the decisions to keep moving in that direction. Or rather that's complete garbage and you need to start over. Um, and so this is, this is the process all along the way of working in a regime like this. Just trying to be very mindful of, okay, are we on the right path, is it working? And that over time you start to see that, um, you know, you go from 30% to really starting to see the signals that you're hoping to see, um, that are very convincing that the

Speaker A: models are working for other AI, for biology companies that are watching this deal and watching noetic success. Could you talk a little bit about any advice you might have, uh, what conditions made this deal possible? I think you've already touched on this a little bit and you know, how could this potentially be reproduced by others?

Speaker B: Um, yeah, I mean it's always hard. I mean one business development in this space is hard because it's very much about, you know, there are a lot of stakeholders, there are a lot of, you know, these are very large companies. You can talk to, you know, hundreds of people within any one company. But I think ultimately it comes down to the science and proving that, you know, one, you are solving an important scientific problem for the partner and aligning ah, along the lines that your, your capability actually does solve that problem. So you know, it's a little bit cliche but I think ultimately comes back to like, okay, what, what is the thing you're building? Does it, is it scientifically working? Um, and then you know, finding the right partner that um, you know, sort of understands, you know, the, the use case and that is really excited about what you're building. Um, and oftentimes that can take time. You know, it's often it's, you know, one that one of the challenges just in startups generally for founders generally, I think Paul Graham just posted something on X about this. Um, is like communicating what you're doing in a clear way where it connects. I think, you know, when you're in like, let's say you're at JP Morgan and you have a dozen meetings with different companies, um, you really need to figure out like what is the messaging that's going to Help people understand exactly what you're doing. Because if on, ah, if you're on the other side of the table, if you're, you know, let's say at a big company and you're talking to 20 different startups, um, you know, you just don't have, you're taking a bunch of meetings, you don't have, um, uh, like the patience to, to like, try and connect all the dots if it's not clear. So I think it's really important to think about, well, am I communicating the things that I'm building in a way that is resonating with the folks that I'm talking to?

Speaker A: I think this transitions nicely to maybe talking a little bit about your philosophy on, um, company creation. Some advice and some learnings from your time at Noetic. You've written about, in interviews you've given, you've written about a specific kind of resilience that's required when building something genuinely new, especially when many people believe very confidently that the status quo is correct. So how do you tell the difference between people think I'm wrong because I'm early and people think I'm wrong because maybe I'm not heading in the right direction?

Speaker B: I don't know. You can't, you can't listen to people that's. Don't listen to people. Like we were, you know, I think for, for many years, uh, you know, people were like, what is this weird company, Noetic? What are they doing? Well, like they're generating data on patient samples. Like, why, why would you do that? Like, there's no perturbations. Um, you know, we, you know, from the very beginning, the models we've been training were world models. People are like, what are world models? Like, some of our investors would be like, I don't know why you're talking about world models. No one knows what this is. Um, and, you know, now, now, you know, three and a half years later, it's like, okay, as of two months ago, everyone is talking about world models. For example, um, you know, and you know, once we, once we were able to, uh, you know, announce that they will, tsk, everyone's like, oh, this is very exciting. Anyway, so I think, I think you just kind of have to ignore the, you know, the spade, the noise and kind of focus in on what you're doing. Um, and you know, again, just focus on the science. And if you're building something that is like, incredibly powerful and useful, then, um, the only thing that's going to like, unlock that is getting traction at the End of the day. Um, so, you know, we try to just focus on traction. And we've, ah, we were just talking about this, um, the other day. We're about to put a blog out on some other capabilities and we're like, okay, well, um, are people going to understand this? And we try to make sure we get things at a level that people can engage with them. Um, but also on some level our work is deeply technical. Um, and so whenever we put something out, it's deeply technical. And the models we're building are not things that everyone are, that a lot of people are thinking about. And you know, it turns out that we end up getting a lot of value from folks that, that are more, are more technical and are, you know, more interested in sort of thinking about things that are, you know, maybe outside of the regime of, of, you know, a lot of other capabilities. So I don't know, on some, on some level you just have to sort of embrace what you're building and embrace who you are and, and keep like, breaking down barriers, I guess.

Speaker A: How did you think about picking your investor that you were building? And what do you wish more investors understood about what you're building?

Speaker B: Uh, yeah. Interesting. So, um, maybe this is a unique M quality that. I don't know if everyone will agree with this. Uh, I really like a lot of our investors because they feel like they're part of the team quite often. And so they are, they are often thinking about what we're doing. They're engaging in the ecosystem. You know, we interact a lot. Um, I enjoy that personally. I enjoy kind of working closely with people and feeling like, okay, our, our board is not necessarily a governance committee, um, only. But they are also, you know, part of the team. They're part of, you know, setting strategy and they're excited about what they're, what we're doing. They come up with ideas, they text me. So, um, I just like that level of engagement. So I think, think, you know, that's not for everyone. I think, you know, maybe founders oftentimes, you know, prefer to have people pretty hands off. So, um, I think it's important to be realistic. You know, what type of engagement and interactions do you want to have with, with your board and with your investors? Um, I also think it's important for your investors to understand what you're doing. I, I think that's, you know, that's not necessarily an obvious thing. Um, but when your investors understand what you're doing, they can actually be a lot more, um, you know, both in terms of Supporting your next financing and um, getting you access to other investors. Um, but also they're excited about the company, um, and they're helping connect you to other companies to help open up opportunities. And I think it's really important for them to really kind of connect with what you're doing.

Speaker A: Noetic really sits at the intersection between deep biology and deep computation. So you need both world class biologists and ML engineers, uh, to talent pools that often don't speak the same language. So how do you think about building that team? What do you prioritize when recruiting?

Speaker B: Yeah, we've been really lucky where you know, we've, the team has kind of recruited itself which is a little bit strange. This has not been my experience. Generally we haven't put out uh, you know, that many uh, uh, job racks. But um, in general, um, in terms of the team structure. So what we're doing on the, on the research side, in terms of the AI research, um, is pretty difficult. So um, we have folks, um, really across the spectrum from pure AI research to uh, ML engineering to data science, to computational biology. So a lot of deeply technical teams, um, and in building the company have just tried to focus on bringing in the sets of skill sets that we need, um, as we need them.

Speaker A: Them.

Speaker B: Um, and so for example, the first folks we hired were software engineers. Um, why did we hire software engineers? Well, we knew that this data was going to be like massive and very difficult to work with and we're going to need to have really great data infrastructure. So we hired our first engineer, our first hires, our software engineers. They started working on the infrastructure. Then we hired the lab, the lab started generating the data. Once we had data coming off of machines, we had the infrastructure. Um, and then all of a sudden we had, you know, some data problems that we needed data scientists for. So for example, aligning the images, you know, data cleanup, data curation. So then we started to hire data scientists. Um, it really wasn't until uh, you know, about two years later, um, that we started building the AI research team because we didn't have the scale of data to actually, you know, train large models. Um, and then as we started training, you know, these large models, all of a sudden we need engineers to uh, help us, you know, scale up model training. Um, and so that's generally how I try to think about it. Um, you know, you don't, you want to, you want to get, you know, as you're building a company, it's good to get people in. Um, as you're sort of arriving at you know, the, the key use cases that you know they're, they're really going to jump in and help solve.

Speaker A: Thank you. Yeah, very helpful. So if Noetics succeeds, what does that look like for a cancer patient sitting in a clinic?

Speaker B: Well, 95 of 95% of drugs will not fail. So that, that's the first, you know, first impact we aim to make is that you know, we will have a lot more medicines for patients and the way we'll get there um, is that you know, we can begin to pos position, you know, a particular type of drug, um, in the patients that are going to respond to that drug. Um, you know, how do we do that using our platform. So if we go very far into the future, uh, today we're in a regime where um, and we're moderately m successful here, um, not fully, but we are in a regime where we are trying to use let's say genomics, um, tumor genotypes, um, to figure out which drugs work in which patients. Um, we're not even sequencing all of patient tumors at this point really. We're sequencing know tumors where you know, we think there's some drug that's going to make an impact. That's a huge miss because we're not collecting data um, to help us develop drugs for those other tumors, um, where we can potentially identify therapeutics. But that's an aside. Um, so really what we're building here is the next version of that. So a, a tumor mutation is a very low dimensional um, profile of a tumor. It's um, a single gene and that a single gene mutation is about the limit of what we as humans can handle in making therapeutics decisions. If all of a sudden you have to consider, let's say two genes mutations or three genes mutations, you know, all of a sudden that level of complexity is just too much for, for humans to figure out. You know, how do I solve this matching problem of you know, a drug versus you know, three different potential mutations? Um, so really what we're envisioning is you know, a world where the models can solve these more complex problems by you know, characterizing patient biology in a self supervised way. Um, and then we can essentially use the model as the decision point around which patient will receive, you know, which therapeutic, you know, initially that could be for you know, a small set of drugs. But ultimately that should be, you know, how a patient, um, and how a doctor decides to prescribe a therapeutic like um, you know, a patient sample from a tumor is taken, uh, that tumor is profiled, um, and Then the model can be the decision framework around. You know, this is, this first set, this is the first set of drugs they should get. Um, and actually you can progressively adjust therapeutics over time, potentially even as well.

Speaker A: How important do you feel interpretability is for these models given what you just brought up?

Speaker B: Uh, yeah, I'm very torn on interpretability. So we have deep interpretability because all of our data layers are labels. So we have 19,000 genes we can define, you know, we have, our virtual cell model defines biology in the context of individual cells, T cells, B cells, macrophages, tumor cells. So what's interesting is we have complete interpretability all the way down. Um, personally I'm, I'm not an interpretability maximalist. I, I would, you know, my personal opinion is, um, I'm quite happy with, with uh, algorithms being black boxed. Um, but I know everyone doesn't agree with that standpoint. And so, um, we have built this in a way that we have full interpretability. And so whilst we can say, okay, um, we can take this pathology specimen and um, the pathology specimen can be used to train a model that tells you whether a patient responds just from the pathology. We can actually go back from the pathology specimen and ask, okay, given the sets of patients that responded, what are all the biological features and mechanistic understandings of these tumors as well from the underlying data layers?

Speaker A: What do you think are some regulatory or clinical infrastructure bottlenecks that the field isn't thinking enough about right now?

Speaker B: Um, it's interesting, I actually think we're in, uh, an interesting place right now with regulatory because, uh, more recently the FDA has been very optimistic about AI applications, um, in clinical development. Um, they've been very optimistic about, you know, beginning to, you know, beginning to move to decision making that starts with, you know, patient samples, not with animal models, not with, you know, in vitro models, but really, you know, decision making that, that starts, you know, directly from patient biology and being, and being in a very patient centric way. So I think it's interesting right now we're in this regime where I think we can start to think more creatively about how we um, can move some of these approaches through the regulatory frameworks. Um, historically people have been, you know, more cautious and more conservative, um, less optimistic, ah, about regulatory challenges here. But I think, you know, we've already seen a couple different examples of, um, AI based diagnostics, um, um, that have been approved in the clinic. Um, I think we're, we have multiple companies that are working on these types of applications. So I think we'll start to see more. Um, so generally I'm pretty optimistic that uh, you know, as long as, and you know, this is my experience with uh, you know, regulatory recursion as well. As long as you're bringing good science to the agency, you know, people are very excited to hear what you have to say and very open to it generally. Um, and I don't think it's such a huge hurdle, uh, necessarily. Knock on wood.

Speaker A: And so we've talked a lot about oncology, but at its core, what noetic is asking and answering is whether AI can learn to model complex human biology from the right data. So what becomes possible beyond oncology and what new doors does it open and what is sort of the forward looking direction there?

Speaker B: Yeah, I mean oncology is huge. There's a lot to do, there's plenty to. Our team is incredibly passionate about oncology. Um, but as you said, complex biology goes far beyond oncology and there's much more biology to discover in other areas. So we're not spending a huge amount of time in other areas today. But there are some areas that we're um, you starting to think about. You know, for example, we've been very interested in liver biology. Um, you know, there are, you know, many different applications, uh, that you could think of. Um, where you know, liver diseases are. You know, there's liver cancers and you know, there are diseases like cirrhosis that progress to liver cancers. Um, and you know, these are very challenging, uh, diseases with like high morbidity mortality. Um, but there are also, you know, metabolic diseases of the liver and you know, a lot of different subtypes of liver disease that are really actually hard to characterize, um, without these types of methods. So I think that's a really nice space for us to think about. We're also thinking about some adjacencies to, you know, some of the tumors that we have already. So for example, um, turns out we have a lot of lung samples, um, you know, in collecting lung samples, um, and you know, quote unquote, normal samples, we've collected, you know, some samples that are fibrotic lung. And so it turns out we have a whole bunch of, you know, lung fibrosis. Just in the context of generating, you know, lung cancer data, um, from so called normal samples, um, and then in the colon cancer sample, uh, set that we've started to generate, you will start to see inflammatory bowel diseases as well. So starting to think about the adjacencies to some of the data that we have already, um, that will allow our models to generalize beyond oncology.

Speaker A: Thank you so much, Ron, for joining us today.

Speaker B: Yeah, thanks, Rachel. This is one Sam.