How CytoReason is Bridging the Data Insight Gap to Accelerate Healthcare Breakthroughs - Ep. 276

0:10Hello, and welcome to the NVIDIA AI podcast. I'm your host, Noah Kravitz. You've heard of language models, video models, reasoning models, and foundational models. And here on the podcast, we've talked a lot about healthcare-specific AI models for things like protein structure prediction. Well, today, we're exploring disease models. The cytoreason disease model is a comprehensive model of human diseases that models and compares treatments in patient groups, helping researchers of all levels make data-driven decisions across the drug development lifecycle. That brief description doesn't really do justice to what disease models and Cytoreason as a company are all about, but our guest is here to help.

0:49Shai Shen-Or is co-founder and chief scientist at Cytoreason and professor of systems immunology and precision medicine at the Technion, the Israel Institute of Technology. Shai is here to tell us all about Cytoreason, how it got started, why the technology is so important and what they're trying to do. and we're grateful to have you here. So Shai, welcome and thanks for joining the AI podcast. Oh, thanks, Noah. Pleasure to be here. Thank you for inviting me to speak about my favorite subject, I guess. Please, and just take it right from there. I don't even need a more pointed question. Tell us about your favorite subject.

1:25Maybe start with your background a little bit and then tell us how Cite or Reason came to be and what it's all about. Sure. So yeah, I'll go back, I guess. at this point, I can think of myself a bit of a dinosaur in the face of doing computational biology data science. I started back in the, I guess, late 20th century, as they say, with the idea where basically, you know, we're just starting, the human genome is getting sequenced and the realization that biology is making a leap from a one tube, one result type field to one tube, a million results type field. And suddenly there's room for, you know, what evolved to what I think now we think about as data science and AI in the context of medicine and life sciences and healthcare.

2:15And for me, that discovery of falling in love in biology, you can do, I was kind of doing a lot of stuff around AI in the late 90s, as they say, very different space. Eons ago. But discovering that you can actually use the same kind of the same type of thinking, but in a space such as life sciences and healthcare was to me profound and kind of changed my life course, realizing that in this space, actually not only, you know, is the data interesting and there's a good, I think, humanitarian cause, but also are the AI challenges are profound because this data is, is I often call it deep more than big, right?

3:00I have, you know, a big experiment is a million measurements on 100 people, right? And so there's way more, you know, way more features. P is greater, greater than N type problem. And that brings really interesting problems in how do you build machine learning models that actually overcome this when there's not that much of a repeat kind of information to learn from. And that brings in a lot of prior knowledge and we'll get to talking about that, I guess, later. Yeah, so that's kind of how I came from this, systems immunology. I'm a faculty member at the Technion. And I realized, I guess, in the 20 years I've been doing this type of work, I realized that as biology, we're kind of fascinated with that two 1 million results.

3:43I realized that we're actually in this amazing times where data is exploding, but value, insight from it, does not explode in the same rate, right? The gap between data and insight, actually, you can think about it like data is exponential, insight is linear. Every day, every day, percent data utilized to give insight is lower. And the question is, how do you overcome this and how do you develop these techniques to ultimately bridge what I call the data insight gap? And whereas biologists, molecular biologists have been investing a huge amount in making this amazing tools that can measure basically now every layer of human biology comprehensively.

4:27The analytical side of this and the AI solutions for this have been missing. The field is still largely a manual field where you give people some data. They sit in front of their computer. You know, they try to figure out, they make some value and insight for this. And I figured that's not a sustainable solution. This field needs to move to ultimately build much larger integrative solutions that bring in many different angles of machine learning, AI statistics and so forth to ultimately bridge this. And it needs to be done in a way that's ultimately reproducible and productized. And that's kind of what launched Ciderys.

5:05So we founded Cytoreason in 2016 with the aim of basically building a pharma AI company that is not a biotech company, that does not develop drugs. It develops an analytical platform, an integrated AI solution to bridge the data inside gap. So that's, I guess, our origin story. Got it. And so then if you're not, if Cytoreason isn't a biotech company, do you serve biotech companies? Who are some of the customers, or maybe if that's not the right way to get into it, tell us a little bit about what Cytoreason offers. So I think it's actually a great place to get into it because this data insight gap exists throughout life sciences.

5:48Everywhere now, biology is a big or deep data field, as I said. And the question now is, you know, OK, well, you know, where is it going to be the matter the most to bridge the gap between the data and the insight? and nowhere is it more important or cost effective and I think ultimately brings the right utility to humanity than to close it in drug development. I don't know if you're familiar with the numbers, they're horrendous, right? If a drug today costs$2.5 billion to develop, most of that cost is actually, it needs to overcome the failures. It's a failure business. Most drugs that you try to develop, even if from the first drug that you put into a human, the likelihood of actually it failing is 90 % to ultimately not making it, right?

6:35And if you're talking about many subdrug, if you're a pharma company and you need to be developing and you have many different assets that's being developed, well, you need some kind of a scalable solution to ensure that your success rate over time grows. And it's a no-brainer that it needs to be a data-driven solution. So CiderEason customers are some of the world's largest pharma companies. Pfizer, Sanofi are examples of companies that Cytoreason has longstanding relationships with. But the same problems that happen with Pfizer, you know, who's developing hundreds of molecules and legs, it happens in a biotech.

7:12So anybody who's developing a drug, and I would even argue diagnostics and so forth, has this problem of how do I make decisions that are data-driven at scale. Right. And so do the Cytoreason models allow researchers, pharmacists, to predict the effects of a drug they're working on? How does that work? Just kind of in lay terms. Sure. So, you know, in terms of the user base, it's really the SatterAid platform is an enterprise solution. You know, we're trying to address the needs of data scientists whose work, because of the data inside gap, keeps growing like crazy. Right. And as you said, you're working with some of the biggest pharma companies and institutions in the world.

7:57So this is a lot of data, big gap, getting bigger. Correct. Correct. And so what a data scientist needed to do, you know, two, three years ago in a pharma company, it's like keeps growing. It's like TEDx because the data keeps exploding and nobody wants to make a decision. Yeah. Having just suffered from the problem that they didn't have time to take the most appropriate data set and analyze it and figure out what it is that they need to do it. Right. So it's data scientists. It's biologists who are not necessarily programming, though this is becoming less of an issue, I think, now. Yeah. You can touch on that.

8:30but who are ultimately driving their particular drug programs and need to make decisions around those in the context of the competition, the standard of care, what other pharma are doing. And it then goes up to, you know, heads of therapeutic areas who need to choose what is the, you know, not only do I want to develop this drug, what's the right disease to go after? Well, you know, there's many diseases. They're only going to give me so many shots on goal to fail or succeed. I need to make those choices, right? Some of those considerations are commercial, but many of them are scientific. And that's what Cider Reason brings to the table.

9:08And it goes on and on in that space. You can think about portfolio management, people who make strategic decisions. Those are the user base. And Cider Reason basically brings in all the world's molecular data in humans right now, integrates it into a single model that allows us to learn from this and ultimately support decisions, use cases, such as how do I prioritize a target? Which target or combination of targets I prioritize? Which diseases do I prioritize for my next trial? Or what subpopulations should we be excluding or including from the trial? Because they'll succeed. So those are really expensive decisions, complicated ones.

9:50And we basically try to bring a yardstick to all the science, the molecular science that's out there. Right. It's a big yardstick. I want to get to how you decided to start building agentic workflows. And if there was something specific, specific kind of challenge, whether on the science end of things or in wrangling different types of data and that kind of thing. But maybe walk us through a little bit kind of how you architected, how you built Cider Reason. And then, you know, describe now the agentic workflows you're using and go into why a little bit. Yeah, sure. I mean, I think in some ways I already gave a bit of a clue.

10:23because I described the data inside gap, right? So imagine you live in a field and I'll give you an example just to make this kind of real. If you're talking about the scientific literature, my field, I mentioned in the beginning, I'm a systems immunologist. I work on the immune system. In immunology, every two minutes, a new paper comes out. Wow, okay. It's kind of humbling, right? And even if you're going to argue that half of them are not worth reading, you're still going to run out of time. Now, similarly, if you're talking about single cell data, RNA-seq data, gene expression, proteomics, things that our audience may be familiar with or not, it doesn't matter.

11:06That data is like while I'm sitting here working, while we're talking, there's data coming out that could be very valuable to drive my decision. So Cytareason as a company is by its own definition of its goal and vision needs to somehow beat this exponential growth in data, right? So it immediately says, I say this to every employee at Sider even, you know, say 80 % of your time you spend on whatever your job is, 20 % you have to spend on how do I make my job obsolete and automated so because I have the next challenge to do. because that data, if we don't, we're going to be beaten by the avalanche of data coming in.

11:51So that, if you think about this as a pitch for agentic AI, it doesn't get better than that. I'm seeing a commercial in my head with agents and doctor scrubs and track shoes and just running as fast as they can to stay ahead of the data just piling up. But yeah, that's, yeah. Exactly. So you basically are constantly in a game in which you need to make it faster. It's actually what's called an evolution. And remember Alice in Wonderland, the Red Queen? Sure. Right? Where she said to Alice, you have to run just to stay in place. Yeah. Yep. Right? And it's also an evolutionary principle of how viruses in the immune system combat.

12:33This is another topic. I can talk to you about this another time. But the red-green effect, so this need for us to continuously run is a huge driver for automation, acceleration, and I would even say the cognitive meta-analysis that we as humans need to do to somehow describe to a machine how we make decisions so that we can automate them. Right. And so with that in mind, I think almost side of reason had the thought that we need a genetic AI even before a genetic AI was there. Right, yeah. And of course, when it came around, we jumped on the bandwagon. Yep, yep, yep. And so it starts at the earliest stages.

13:12I need to bring the data in, right? To bring the data in, you could go the manual route, which is like to have people bring it in one by one, totally unsustainable given the data keeps growing. A paper every four minutes if we include the bad half. Yeah, that's a lot of data. Or, you know, data sets and so forth. At every molecular level. So you really cannot do this manually and strive to get that level, right? You can build pipelines that you want to process this. And as soon as you do this more and more with kind of molecular level data, you realize there's, in biology, has a lot of these exceptions and outliers and so forth.

13:51So ultimately, a more appropriate solution is to teach a machine a workflow that may be very complicated, where humans make decisions, but you can see it, and then start that automating process. Now, would it work perfectly from day one? Depends on the complexity of the data that you're going to. But if you then, you know, you put a QC process that you start with manually, and then you make that automated as well, and so forth, you can build processes that really accelerate your data intake. And that's just the most obvious place where the agendic AI comes in, right? It can come in other places around, you know, decision supports.

14:27We'll talk about this one. So thinking about keeping up with the data, the literature in particular, were there specific techniques or, you know, there are obvious advantages we've talked about, just agents being able to go out and do the research and grab the data kind of obviously is a, you know, a game changer. But other things that you discovered about working with agents to curate and review the medical literature in particular that jump out at you? Well, I think it's a wonderful question. I'll answer, I think, on two angles. So first of all, just to say why in such a data-rich field somebody needs literature.

15:08What you can think about it is data. You can say, well, that's data, right? But I actually want to uniquely identify that data from other data because I would argue that literature is already at a stage of knowledge. And biology has a lot of data that isn't yet knowledge. Ciderism deals with all the data, but also the literature. And the reason we need it is because of, well, A, people want the knowledge, right? But the other side, which is more interesting, is when I describe that this is a deep data field where there's way more features than there are kind of measurements or samples and so forth, the way that people make decisions, you basically cannot just stick this into, you know, a machine learning model.

15:50And it's going to be basically an overfit, right? And the way you kind of deal with this is actually by the integration of prior data, which comes from the literature, that allows you to narrow down the search space in the bar you feel. Right. Okay. It has two advantages. One advantage is that you do make, it's easier for you to make discoveries. And there's another advantage, which is relates to our customer base. And I think in general, how people make decisions, large decisions in the face of uncertainty, which is that they want to stand on the shoulders of giants, or at least on stand on some level of confidence, right?

16:26So being able to connect new novel discoveries, emerging phenomenas and so forth that the AI model produced to knowledge that I solidly believe in a firm is actually an important thing for our customer base and for any scientist to actually make the leap, right? Because it's gonna, the next stage could be, you know, it usually would be an experiment. it can sometimes be a very expensive experiment. And people, it's not enough just to have a predictive model. People are seeking from our, our customers are seeking from us and where we strive for it to be a mechanistic model. Explain to me why that prediction makes sense.

17:08And give me trust in it. And the literature brings that piece in, right? From an agendic and I perspective, that means also, for instance, as an example, confidence scores on the literature are a really key thing for us, right? Because this literature is complicated. There's not a huge amount of instances of any one event being, how sure are we that this particular kind of description of a biological event is actually correct? And that for us was a huge piece of entering of how we kind of been pushing the LLMs within CIDREs and then the Agenda KI workflows. And it goes, I kind of mentioned, it goes everywhere here.

17:47We need to be really sure, you know, Gen AI is awesome, but we need to have the high quality. And so we've been putting a lot of these guardrails, if you like. Yeah. Where do the confidence scores come from? Are you, is CIDR a reason generating them? Are they in the literature? So you basically can come up with a variety of different techniques in which by sampling the literature, right, and also fitting, you can, you know, by sampling the literature, you can build that confidence on one hand by putting an LLM RAG component, right? So you're actually doing retrieval augmented generation and kind of querying this to be more certain about what it is that I'm looking for.

18:24Right. All of those, and there's a variety of other techniques. There's also the kind of the, what we call biological expectations or biocredibility in the end to check ourselves on this. And so that it's a loop that keeps improving. All of those are techniques that allow us to basically build the confidence that we need for these heavy decisions. On one hand, the necessity to leverage an AI and a genetic guide to basically move forward and do this on large scale. And on the other hand, to ensure the confidence is high. I'm speaking with Shai Shenor. Shai is co-founder and chief scientist at Cytoreason, the company we've been talking about.

18:59And he's also professor of systems immunology and precision medicine at the Technion, Israel Institute of Technology. We've been talking about cytoreason and just recently the importance of building trust in the model's output. And, you know, it's something that applies to generative AI in any situation. But as you were saying, Shai, in biology, in precision medicine and drug discovery and pharma and everything, these decisions are both, you know, literally can be life and death for lots of people, but also quite expensive and involved, you know, saying go involves a lot of resources being put to use, a lot of money being spent.

19:35And it made me think, Shai, we've had a couple of conversations on the podcast with folks in the protein sequence prediction and generation space and other drug discovery related spaces. And I'm wondering about cytoreason's place in the workflow, in the researcher's workflow or the end user, whoever's using it. And, you know, when you mentioned about making these decisions and experiments being expensive, I've talked to folks before. I've read about folks using AI models, generative AI, to do sort of simulated experiments, right, before moving to the wet lab, being able to run and kind of narrow down which ones are worth the cost and the effort to do.

20:18Are your customers using Cider Reason kind of in a same way or what does a workflow look like? And then out of that, I wanted to ask you if there's a time you can share with us where Cytoreason's workflow enabled something, you know, really unique from an end user. So can you talk a little bit about the workflow, if you could? Sure. So from a workflow perspective, there's maybe two points to say. And it sounds like you have been talking to some interesting people doing interesting stuff around ProteinSar. So I'll differentiate ourselves from it. So Cytoreason is a company, and this is also interesting, I think almost from the NVIDIA kind of marketplace and kind of the company, the other reason is a company that NVIDIA invested in.

21:01And I think we stand out uniquely within those. Because if I look at the healthcare flow, right, there's the chemistry of it, right? There is the biology of it, and there's the kind of clinical side of it. Okay, right. And if you look where AI has been playing a big role at this point, It's certainly been in the chemistry space, chemical structure, and I would put protein structures there as well. From small molecules libraries to protein structures, there's a huge amount that's happening with kind of, you know, NVIDIA GPUs and, you know, generative AI and so forth to basically build those molecules.

21:37And of course, anything that gets built there, there's a simulation test, but ultimately somebody puts in experimental tests. And experimental tests are usually, I would say, at early stages, they are what would be called an in vitro experiment. There's no animal, there's no human, you're just texting to see, well, okay, was this antibody that I just kind of simulated when I generated, does it hold the properties and can it be a good direction? On the clinical side, there's also a lot of, I think, agentic AI happening with a lot of kind of shortening, say, kind of recruitment for clinical trials and so forth.

22:17There's a lot of space happening in the electronic and medical record. It's a relatively well-defined space. There's a huge amount of almost like human operations that goes there. And then I think agentic AI is playing a big role in. Side of reason is quite unique in that we're focused on the biology side of things. And biology, if you compare them to the two, is actually the biggest unsolved problem. I would say today, if I look at pharma, the two big problems from a science perspective, are one is that we don't have a good understanding of the biology. And you see it in clinical trials that phase two, which is the first time we tested in humans, is where the biggest failure rates are.

22:57Right. Okay. Okay. So it tells you. And then the second piece is the human diversity. So the biology can be, you know, you and I may not have, you know, we may have the same indication and so forth. We actually may look very different and could be for different causes. We still don't have a good understanding of this. That's where insider reasons are playing. And bringing AI there is, you know, the search base is way, way, way bigger than the chemists. And so it's an early stage to build on, but it's clearly biggest problem. And that I think where we'll see companies going on. And certainly that's kind of where we've been kind of leading and NVIDIA kind of putting the, or I think our trust in us has been a huge thing for us.

23:39So, you know, I think if I look at that space, are the users behaving there? Well, first of all, they need to explore the disease biology and then they need to think about their use cases. And again, the use cases is what would be a good target to choose from, given I need to have it work in this particular disease. and given that I know, you know, this is, you know, this disease already has a bunch of standard cares that I need to beat. Right. And I know that there's people who are not responding. And what is it that's about them that maybe I can target? And there's other companies that are developing and there may be, you know, out to market before me.

24:16So they need to, all of this commercial questions need to come into a scientist thinking about disease biology and saying, where's the niche that I can come in? And so whether it's target prioritization or bigger than this indication, choose the next clinical trial. Is it happening in RA? Is it happening in Crohn's disease? Is it happening in Alzheimer's? That's not an easy, those are the use cases. And so if you think about what side reason brings, it tells this particular target is the best priority to go for this disease. Here's a bunch of mechanisms why we think this is the case. And the users can go and do small tests.

24:55It's very different from the protein structure ones I mentioned before. But small tests that actually validate the top hypotheses, build confidence in the AI prediction, and then you go and execute on them. What's the feedback been like from your users? And I'm wondering, I mean, go anywhere you want with this. I'm wondering if there are certain areas that have been brought to your attention to focus in on whether it's that the users have been kind of poking at a certain area and wanting more functionality, or if maybe something you didn't expect popped up and it's a different path to look at.

25:31Sure. So I think in general, so it's a very interesting question. There's a lot of these, right? So in general, the users want a lot. Yeah, no, right. What does any user want right now? Yeah. But I'll just mention a few directions that you'll just see how they themselves struggle, right? So on one hand, you could think about this as I'm invested in a particular asset. I just paid or I invested a huge amount of hundreds of millions of dollars to manufacture a drug. And what I want to do is deepen. I want to study that, get every possible layer and model every possible layer here that my prediction is the best.

26:08And on the other hand, orthogonal to this is you can say, and this is obviously like a person who's a program lead for that drug. If you ask him, that's what he wants to do. Then you go to somebody who's in charge of an AI strategy for a pharma company or is the head of a therapeutic area. And they say, well, that's one drug. Obviously, I care about it. but I have 100 drugs that I actually am simultaneously developing, and I need to evaluate them across tens, if not sometimes hundreds of diseases. We need scale, right? Those two are orthogonal, and you need to basically care to do it to both, because the science is always in the depth.