BioCentury This Week

Ep. 323 - Agentic AI: From New Targets to the Clinic

• BioCentury • Season 6 • Episode 323

Share episode

AI is bringing sweeping changes to drug development, from how targets are discovered to optimizing clinical trials to maximize an asset’s chance for success. On a special edition of the BioCentury This Week podcast, IQVIA’s Greg Lever joins BioCentury’s analysts to discuss agentic AI’s short- and long-term prospects to help biotechs discover new targets, predict success in preclinical development, and enhance clinical operations. This episode of BioCentury This Week is sponsored by IQVIA Biotech.

View full story: https://www.biocentury.com/article/657086

#Biotech #Biopharma #DrugDevelopment #ClinicalTrials #TargetDiscovery #AgenticAI #GraphRAG #DeRisking

00:01 - Sponsor Message: IQVIA Biotech
01:22 - AI in Biotech
05:01 - Machine Learning
06:21 - Generative AI and Language Models
08:37 - Agentic AI
12:43 - AI in Target Discovery
23:44 - AI in Clinical Trial Design

To submit a question to BioCentury’s editors, email the BioCentury This Week team at podcasts@biocentury.com.

Reach us by sending a text

[AI-generated transcript.]

Eric Pierce:

BioCentury This Week is brought to you by IQVIA Biotech. For biotech companies striving to bring innovative therapies to market and maximize patient impact, IQVIA Biotech is the trusted CRO of choice. Backed by 25 years of unparalleled experience and deep therapeutic expertise, our full-service clinical development solutions are purpose-built to accelerate success. IQVIA Biotech helps early-stage biotechs de-risk by developing strategic clinical development plans, guiding drug candidates along the most promising pathways. Leverage data-driven models and dynamic tools to craft a compelling value story and maintain momentum through every phase of drug development.

Jeff Cranmer:

AI is the buzzword everywhere these days and biotech is no stranger to that. Welcome to a special edition of the BioCentury This Week podcast. Today I'm very pleased. we have a special guest from our sponsor, Greg Lever, he's the Director of AI Solutions Delivery at IQVIA. And joining me as well are my colleagues and podcast, regulars, Selina Koch and Lauren Martz. And we're gonna cut right to the chase, Selina, when you think of AI in biotech, uh, what does it say to you?

Selina Koch:

right now if I try to take a step back and like take in all these changes that are sweeping through the industry so rapidly, it kind of makes my head spin a little bit. conversations about, how to use AI and agentic AI and drug development, are now, you know, being happening in, so it seems like in every single company, at every investment firm, and it's even. Maybe even hard to go to a meeting where the topic doesn't come up right? So things are changing at kind of dizzying speed, but because of that, we're in this like moment where there's this wide variability across the industry in terms of how much or little people understand about the technology and how proactive they are in engaging with it, I think. And so I think it's a really nice time to have this conversation to kind of take stock. Of where agentic AI is, making some practical inroads and, where it has, other applications in the future. if I could say one more thing, I don't think that's always obvious, like where AI is gonna make contribution. like I don't think many people would've guessed that writing code would be among the first breakthrough applications of AI that. First jobs under threat would be programming jobs, for example. But you know, in retrospect it makes sense when you think about the data sets that are available in that space. There are, well, I don't know, or Greg might know this, um, tens of millions of programmers who regularly check code into GitHub. So there's this really big repository that not only has a lot of code, but has like a code progression, leading to code working or not working. So you have. Also this reference to ground truth. but when I think about biology, it's really not that straightforward, right? We have this like fragmented evidence ecosystem. We have a few large scale data sets would have that kind of reference to ground truth. that makes me question in the near term where there's gonna be practical application. Like if you think of protein design, for example. I can imagine that optimization of antibodies and proteins along certain parameters will, will be happening. It is happening, right? But wholesale de novo design could take a while longer. I dunno. If you all disagree, feel free to chime in. But

Jeff Cranmer:

Lauren, do you have any thoughts that you wanna throw out there before we, uh, go? I see Greg nodding along and we'll, we're gonna bring him in in just a minute, but, uh, Lauren.

Lauren Martz:

Yeah, I think we can pass it to Greg momentarily, just to follow up on some of the ways that this technology might be finding its way into our industry and already is. there's protein design, there's target discovery. And that goes all the way through to ways that, AI can support clinical trial design. you know, we've been hearing about this for years. I think a lot of what we've heard initially is that. This can be used to help with site selection, you know, with, finding the locations and all the different dimensions that go into that. This is a place where data can have a big impact. You know, where the patients are, where the right investigators are, um, you know, how many trials are being run in that, in that site. This is something that we, we've already seen this in action, from Greg, I'd love to hear more about how AI can support clinical trial design and everything in this spectrum of drug discovery, through drug development.

Jeff Cranmer:

Alright, well let's bring in Greg. It's Greg Lever, director of AI Solutions Delivery at IQVIA. And Greg, I know, we've chatted a bit before and I think you thought it was important to kind of set some definitions to get us oriented. why don't I hand it over to you?

Greg Lever:

Thank you. Yeah, thanks for having me. It's great to be here. Yeah, and I think to Selina's point, there's a lot going on in the industry and I think it can be quite difficult to get a, a proper grip on a lot of the terminology that's out there and the different approaches and, and also why they matter and, and, and what it is that we can, we can actually, you know, use here. So I think to act as a, a starting point, it can be, helpful to think about. How all the different terms interconnect is specifically aspects like ai, machine learning, generative ai, ai, and if we start with ai, as a broad term, artificial intelligence is really, you know, used to describe all of these sorts of initiatives. And it was, started as an academic discipline in the mid 1950s. and there are now vast different types of approaches and algorithms within this space and. One of these approaches is, machine learning. So this is where, you know, typically you have algorithms that can be used to train a model so that it can learn from data, make predictions based on unseen, but related data. So, as a more concrete example, say I have data on adverse events that occurred in various clinical studies. You can build a machine learning model to predict the probability of adverse events of studies that the model hasn't previously seen. But there are limits here to this in terms of only being able to answer this specific question. It needs very specific data. but what's been really impactful more recently is generative ai. So this is where. The model can generate new outputs, in the form of video, audio, or language like we see with large language models, you know, also referred to as LLMs. And it's these sorts of approaches that are made possible by the work that was done on transformer models. and so, you know, there was a landmark paper from 2017. A year later we saw the first release of the generalized pre-trained transformer or GPT model, and then later iterations of this is what were launched in 2022 in the form of ChatGPT that was, you know, took the public by storm on the internet. And so LLMs like those, powering these chat bots like ChatGPT. What they're doing is they model language as a sequence of word fragments or, or tokens, and generated token by token, it's this new text. That's coming out. It's based on all of the preceding text that has gone into the model. And so with enough training, these statistical dependencies among these tokens, this interconnectedness, it proves sufficient to actually produce what we see as conversational text. And so, you know, often it's basically indistinguishable from that of a human counterpart. the link here for something like clinical development prediction is that in the same way that language model learns from, you know, the grammar, the contextual logic of language from vast bodies of, basically internet scale text data sets, clinical development models based on these same types of approaches can begin to infer patterns of progression. Um, you know, when they're trained on data from things like. Preclinical readouts, clinical studies, approval documents, just to name a few. both types of approaches can recognize these past events, but exploit these dependencies, this interconnectedness to predict future sequences. So that's whether, what's the next word in the sentence? based on a question that's been asked in a prompt in ChatGPT or say the next milestone, in a clinical program. And so when we think about these language models, that there's also another aspect within, within AI that we're hearing a lot about at the moment. and this is agentic ai. And you know, really, really, if we think about a quick definition here, agentic AI is talking about AI systems that can act with agency. And what that means is that they have this autonomous. ability to analyze data, but also make decisions, execute tasks, if that's what you need them to do. So, unlike traditional machine learning or generative AI approaches. These agents, they're designed to reason, but also to plan and act independently on these plans. And basically this gives this whole new level of automation, and adaptability and especially within clinical research. So unlike traditional language-based, applications, I love that I'm talking about traditional language models as if this has been going on for. Decades. Um, and you can see how fast this is moving. So now you can dynamically choose tooling to incorporate maybe reasoning, adapt their analysis based on the situation at hand. And really what differentiates, agents from, from more conventional AI approaches is, is. Really, it's a combination of four things. They've got an an underlying model that they're utilizing. So they might have a language model, they might have a reasoning model, essentially something that serves as like that brain. They also have tools. AI agents have things like maybe, um, databases or, or APIs, other tooling, maybe even conventional machine learning models that they can use, to inform, the insights they're bringing back. But they also have, um, they have memory, they have additional information that's been brought in. And the key thing is, logic that helps the agent figure out what do they do next based on that current state and these decisions. So really to, to finish that up, the analogy to think of with agentic approaches is, A language model can, if you ask it to recommend the best places to visit in Tokyo, it will do that for you. And it do, do it pretty well. an agent can do that as well. But it can also book your flights, your accommodation, make dinner reservations, you know, book you into a show. these are the sort of autonomous aspects that we're then seeing. And you know, as you can imagine, within life sciences, healthcare, clinical trials, there are, there are huge amounts of guardrails and other approaches that we need to make sure are in place before we just, you know, let these agents go off autonomously and do what they want.

Selina Koch:

that was super interesting. so when it comes to the scope of agents, I guess, do you think it's best when an agent is, designed to have a very specific function that it fills and then you string together different agents? With some sort of orchestrator or how are people thinking about building more complex end-to-end systems?

Greg Lever:

Yeah, no, that's exactly right. The the real key advantage that we see with agentic frameworks is that you go beyond this. Initial situation you might have seen, say five, 10 years ago, where to my example, before I've built a machine learning model that can predict adverse events. I have this very specific question and then that's it. I can't go beyond that. Whereas if I, build my very specialized, framework of agents, some which may have access to data, about adverse events, some may have data on, um, preclinical readouts approval documents. But they also have those language models so that yes, they can bring those insights in. And just like you said, Selina, there's, there'll be some orchestrator agent, which not only, brings in the insights from the different agents, but also. Fully understands the intent of the question from the user so that now I'm at a state where, yes, I can predict if there's gonna be an adverse event in an upcoming clinical study, but actually I can go beyond that and answer additional questions that we didn't have to bake in at the start of designing this whole framework. And that's, that's kind of some of the magic and the power that comes out of these things and, and actually how we utilize this in the best way. That's still not a completely solved problem. That's the exciting piece about, agentic frameworks at the moment.

Jeff Cranmer:

Let's bring that to target discovery. What can and can it do in target discovery?

Greg Lever:

No, I think, I think that's really interesting. It's, where these source of approaches, you know, they're currently being applied, in, target discovery in that earlier, piece of the discovery phase. especially when you're looking for. Opportunities for selecting or identifying, um, novel targets. You know, we know that those existing pain points are, how do we, how do we identify those novel targets? We've got all this fragmented and disconnected data. We, we might have limited resources for actually doing a really deep landscape analysis. And actually there's just a huge risk of missing opportunities, and especially maybe in rare or emerging indications. Um, and so there's a couple of approaches that IQVIA uses to support these sorts of challenges. And so really you can think about it, is taking a language model that's been trained on an internet scale dataset. you can think about this as a library and you can think of a standard language model like. Someone who's read this entire library, they can recall what they've memorized during the time they read it. If they're asked a question, but it's based on what they remember, that might be a little bit incomplete or it might be outdated. And so the first step in going beyond this is, is known as what's called, um, retrieval, augmented Generation or RAG. Basically, you can imagine giving person like a librarian to point out here are the really useful parts of the library that are relevant for the question that you have. and, and maybe there might be updated materials that the librarian can bring in that the person didn't have originally available to them. And then the second step is to go beyond this, um, using an approach called GraphRAG. Basically what this does is it creates a knowledge graph so that in practice, um, you know, now imagine the librarian is even more sophisticated. They can, help the person understand these are the important sections of the library, but these are the most relevant books. And also these are the relationships between concepts across these books. Like for example, how, you know, ideas in one chapter might influence another, how different authors discuss the same point from different angles. it allows your language models generate not only, um, informative answers, but ones that are, you know, contextually rich and connected. And so the agentic piece, although also comes in when we use these specialized agents to extract data from different places. It might be scientific literature, trial registries, approval documents to name a, name, a few. We identify these underlying entities within and across data sets. And essentially what it, what it allows you to do is have this optimized language model where now you can start answering questions like, what are the future trends expected in a particular disease area? What are the potential growth areas to consider, within my TA of interest, you know, are there some indications that show. more promise of their pathologies or their mechanisms of action. So really, it's the sort of thing that may enable a biotech to move beyond just, you know, incremental innovation, and actually pursue some, some truly, novel approaches.

Selina Koch:

So on the target discovery front, when it comes to ingesting all of those different kinds of data, making sense of it, like what would be some shorter term benchmarks of success that you would be looking for?

Greg Lever:

And I think this has to really. Align with your metrics for success in terms of your overall R&D pipeline. And so, if that's, if there's, if there's maybe kind of more, more assets that you've got in mind and it's, it's always gonna be these longer timescale, considerations of. Right. This is what I'm looking at to say de-risk my R&D pipeline for the next say, 10 years. If I'm looking at a 2035 roadmap, these are the mechanisms of action or these are the interesting, kind of pathologies or indications that I need to be identifying now. Potentially there's also opportunity to course correct, um, existing R&D pipelines. you might need to optimize in some way, but I think that's where you might think about a different type of approach. maybe if you're thinking about probability of technical and regulatory success.

Jeff Cranmer:

That's fascinating stuff, Greg. I'm curious what, what's next?

Greg Lever:

Exactly, so, so while this, this knowledge graph creation work is the sort of thing that's happening right now as agentic frameworks mature and we begin to expand on their capabilities, you can imagine agents completing tasks like, okay. Go through this knowledge graph that's available to you. Identify indications where an asset may not have been successful. Based on data available. You might be able to utilize existing subpopulation analysis models to predict where patient subtypes may in fact respond well to a previously failed asset. And you know, this, this supports things like defining eligibility criteria for future studies. In a similar way to looking at, you know, approved therapies and identifying potential areas of expansion. And this is something that drug repurposing as a concept has been working on for, for a while, but it's really going to be accelerated by agents being able to work autonomously over this data. But also being creative about how to bring back these insights. And I think that's something that often gets underestimated is the language models capabilities for, for creativity, but also being mindful that that doesn't turn into hallucinations, which

Selina Koch:

That was what I was just gonna ask you about how, how do you be mindful that it doesn't turn into hallucinations? I mean, you wanna use positive controls. I assume, like any experiment does it tell me the things I expect to see, but are there other tips?

Greg Lever:

So I think really one of, one of the things that's, that's crucial to thinking about is in the same way that you have to utilize various metrics and key performance indicators, when you think about any kind of, predictive model, whether you've just put something together in Excel, or you've got a simple machine learning model. And agents have exactly this as well, and so large language models will have things like temperature and other parameters and aspects that you can look at to test. Whether it's hallucinating or whether it is genuinely being creative. And I think one example to think about is you can think of this future where, you might have a, a biotech with an interesting asset or a particular mechanism of action. There might actually be a portfolio of, say, rare diseases with a shared mechanism, which can be impactful over multiple, relatively smaller patient populations and thereby making the asset more attractive to investors. And it's the sort of thing that. Right now is a very manual effort. We'll take a lot of research and a lot of, a lot of work to generate this sort of hypothesis, but it's the sort of aspect that if you have an agent that has this key task, but it's been set off to do, um, it can make it, it can make it much, much more easier to bring back in those sorts of insights.

Selina Koch:

That's a very optimistic example given the, uh, barriers to the business case in rare diseases. so I like that one. Okay. Well, we've talked some about, the applications, you know, near and short term in clinical development and those, for target discovery. Well sandwiched in between those two things. There's designing a molecule, predicting whether or not it's going to be successful. let's dig into that a little bit.

Greg Lever:

Absolutely. So I think one of the things to think about, in terms of some of the forward looking future thinking aspects as well, is that. Once a, a target has been identified, the sorts of things that maybe early stage biotechs can think of is that AI agents are going to increasingly be used to assess things like technical, regulatory, operational success, and really how they can reduce their risk and increase confidence in their, in their development strategy. And so the way, the way these sorts of things will work is there'll be these specialized agents that will be able to understand aspects like, okay, I might have a PTRS agent, which is looking at mechanisms of action, patient populations related, existing approvals. I may also have data retrieval agents looking across scientific literature or conference abstracts. I may have trial search agents looking across existing clinical trials, and so this then gets all brought back together by an orchestrator agent that not only brings in the insights from these specialized agents, but. Can really understand the intent of my user's question, which is basically saying what's the potential of my early stage, asset? And then by extension we can utilize existing molecular design approaches or target design approaches to say, and is there some better design that I could think of here, um, that will like really accelerate my clinical development?

Selina Koch:

So in protein design seems to be a little further ahead of say small molecule design, unless I've misunderstood. Um, but there. This neat idea I heard recently of, so if language models, they tokenize language, as you were saying earlier, where the smallest unit of meaning is the word as opposed to the way we learn language is building it up from letters. that there is an analogy in the small molecule space of tokens or bits of molecules with certain functions and that you might be able to have like. A tokenized library of functional components that could be built into anyway. What, tell U.S. a little bit on the small molecule front, what you're hearing. That could be a be a step change?

Greg Lever:

So I think what we can think of in terms of the protein and antibody modeling space is things like protein language models where that fundamental token or fragment is either an amino acid, if we're thinking about proteins or nucleic acids. If we're thinking about other structures like DNA and otherwise. And I think really. You can think about molecular language models that actually build up this concept. at the atomic level, and potentially also at the, the electron level as well. And that's actually where a lot of my, academic, studies and academic research was in, is in how do we build up a sufficient electron density to understand how these small molecules can best interact. Um, with their targets and being impactful in a clinical setting. And it's something that the language models are definitely going to, to really accelerate, um, in the coming years. And it's a, it's a space that we really need to keep an eye on.

Selina Koch:

That was very cool. And then if we wanna even go further afield, I guess in our imaginations here, what can happen? I just heard a really interesting talk at our Grand Rounds conference on, um, the possibility of quantum computing, really speeding up simulations, particularly around all the electronic states and things like you were just talking about. but that's, not here yet.

Greg Lever:

And that's probably also an entire podcast

Selina Koch:

Yeah. Yeah. Yeah.

Lauren Martz:

I think it would be great if we could get into a little bit about the clinical trial design AI applications, where we are now, and where we could be, you know, in the future?

Greg Lever:

Yeah, that, that's a really interesting point and I think there's a lot to to speak about, but I think what I would say is. AI is really transforming trial design. It's, it's doing things like quantifying site and patient burden. being able to look at things like protocol complexity, but also simulating operational outcomes. So, not only can we use these advanced analytics and other conventional machine learning approaches to really predict how different decisions within the design process, like eligibility criteria and visit schedules. How they might impact recruitment rates, um, patient retention, um, or site performance. But now with, increasingly generative ai, it's enabling these sponsors to, to really optimize these protocols, really say. Okay, this is what may happen in the current state, but how do I make this even better? How do I forecast that site activation? How do I select sites with the highest likelihood of success? And so really it's, it's getting to those pain points of, we know there's a lot of protocol complexity. This is slowing recruitment. We know there's uncertainty in these operational outcomes. and also we know that the biotechs have these resource constraints. And so really what you can do is you can, you can begin to demonstrate this value and sort of leapfrog across a lot of these constraints, by utilizing ai.

Selina Koch:

So looking across these different domains, target discovery, molecule discovery, clinical trial prediction design. When you're advising biotechs today, or if you were to, um, and you had to give just a very short list of like, these are the most useful applications right now across those domains, like what would you say.

Greg Lever:

I think it's about making sure that you've, you've tried to have that coverage across the clinical development life cycle or the, or the pieces that are important for you. to make sure that, okay, in the discovery phase, maybe you are thinking more about target id, generative approaches for design. when you do get that, asset into the clinic, really how are you optimizing, your protocol? And then also maybe when you are thinking, about the regulatory space. how are you utilizing AI to really think about what are the compliance changes that are happening globally or in my target market? And how do I need to basically de-risk, my development and be able to anticipate these things and act earlier? And that's, that's in a nutshell what these AI approaches are providing you the capability of.

Selina Koch:

Is there any application right now where you'd say, just steer clear of that one, or don't expect a lot from that just yet?

Greg Lever:

So I think one of the biggest challenges that we might want to leave to slightly more, well-funded, outfits is the notion of in silico clinical trial simulation, and this is an aspect that has had lots of attention over. Many years from many different types of approaches. It's something that, generative ai, agentic ai, and specifically foundation models are going to have a massive impact in. But it's still very early days.

Jeff Cranmer:

well, we've been talking AI in biotech with Greg Lever of IQVIA and Greg, I'd just like to get your closing thoughts?

Greg Lever:

Yeah, absolutely. I think really one of the things to think about is that for those. Biotechs, early stage biotechs, those sort of outfits. agentic AI really offers the ability to compete at scale. and whether it's identifying novel opportunities, designing smarter trials, you know, accelerating development with, with fewer resources, Not only that, the other approaches that we talked about, be it generative ai and also conventional machine learning, could be helpful in, in attracting investment, you know, utilizing these approaches to give this more objective, data-driven view of asset risk, but also asset value. And so supporting things like capital allocation, the potential to de-risk these investment decisions. I think that's gonna be really impactful.

Jeff Cranmer:

Excellent. Well, Greg, thank you so much for joining us. the way things are moving so quickly, we're gonna have to have you, uh, back on tomorrow. I don't, I don't know. It's, uh, tough to keep up with everything that's going on in, uh, how AI is helping to improve, how we work in biotech. so once again, this has been, uh, the BioCentury analyst team, speaking with Greg Lever, Director of AI Solutions Delivery from IQVIA. a special thanks to IQVIA Biotech our sponsor, as well as Kendall Square Orchestra, the Boston based, ensemble that, does the music for all of BioCentury's podcast Tickets on sale now for their new season. we will catch you on Monday with the regular BioCentury This Week podcast.

Alanna:

BioCentury would like to thank IQVIA Biotech for supporting the BioCentury This Week podcast. To learn more about how IQVIA Biotech can help you turn your vision into venture capital, go to IQVIABiotech.com/visionaries

Podcasts we love

Check out these other fine podcasts recommended by us, not an algorithm.