Cracking the Clinical Trial Code with ConcertAI

Speaker 1:
Welcome to TD Cowen Insights, a space that brings leading thinkers together to share insights and ideas shaping the world around us. Join us as we converse with the top minds who are influencing our global sectors.
Brendan Smith:
Okay, welcome back to another episode of Machine Medicine: AI & Health Care, TD Cowen's podcast series where we bring you the latest and most important takeaways from the state of AI and the healthcare sector today. I'm your host and TD Cowen Healthcare Analyst, Brendan Smith. And today I'm joined by ConcertAI's Vice Chairman of the Board, Jeff Elton. Jeff, it's great to have you and welcome.
Jeff Elton:
Thank you very much. Glad to be here Brendan.
Brendan Smith:
So, for anyone new to our podcast series, Machine Medicine aims to break down the use of artificial intelligence across healthcare into bite-sized digestible points, one episode at a time, highlight the biggest misconceptions, and then recontextualize each piece back into the bigger picture. And today, Jeff and I are looking to dive into how ConcertAI is rewriting the playbook for designing and enrolling clinical trials with AI and data-driven precision and how transformative this approach could really be for the field in the future. So, Jeff, let's just dive right in.
I think it's fair to say, most investors at this point have probably come across AI within the drug discovery realm, in one capacity or another. Though I think, most of the headlines seem to have focused more on target discovery, biosimulation, what have you. But you all at ConcertAI are really reminding us that this technology has use cases across the entire drug development spectrum and in particular how we design clinical trials, as well as decide on and find the right patient populations in which to run them. So, maybe let's just first start with a thousand-foot view of why you think this aspect of drug development is ripe for data-driven efficiencies and how ConcertAI is leveraging AI to really spearhead this movement.
Jeff Elton:
So, Brendan, thank you. And I think it's a good setup. So, there's a few layers to the answer to your question. One is, clinical development processes themselves, and probably the most mundane part is huge amount of administrative processes, a high labor component to it, very sequential processing to doing things. And so, it's not one that's been characterized by efficiency. And in a way, if you think about where generative and agentic AI can really actually kind of has its power and is generally well accepted, it's in areas with that kind of background. So, it's one of the reasons why pharma companies, all their scientific writing has probably already started converting over to generative AI backgrounds. So, one part of it is just to kind of take a lot of what I'll call the non-value-added components and be able to convert it and let AI take over to it.
But then we get into actually more nuanced and arguably a little bit more important in that area. And I'll give you a vignette without mentioning the sponsor, which is there was a trial and a study, it was having problems accruing any U.S. patients. And when we actually used our tools to do an analysis of what was the problem and that it turned out that the control for the U.S. population was not considered part of the current standard of care. It was in Eastern Europe, but it was not in the United States. So, no U.S. physician wanted to have their patients randomized and put on the control part of that particular trial, because it didn't represent the current thinking about how the standard of care would actually be managed for those particular patients. So, they actually were almost getting no subjects to participate. And the problem for the pharma company is, it may not even support a U.S. registration if they don't have any U.S. patients.
So, we're talking about a high cost kind of item on that. Now let's take your question, which actually has about AI drug discovery. So, AI drug discovery, at least when you take the data from the different companies and ones I'm sure you actually have covered and spend time with, processes that may take four or five years, they've been able to kind of benchmark, get that down to 18 months and pretty tremendous. And this idea that I can actually be that much more efficient in running through and generating a lead series. But the problem is when that comes out, it goes through the same system. In fact, again, no disrespect to the drug discovery companies, but most of those time savings will be bit away if the rest of the process doesn't actually have some of the same kind of efficiencies and effectiveness coming out of that.
So, first and foremost, what we really want to be able to do is, simulate that program which they did, let's call AI's own version of a rational kind of process, because they're using lots of infrastructure and alpha fold and few things to kind of progress that. But can we simulate that to help them determine which patient population do they want to focus on for the active treatment arm for that, maybe even go between parts of that lead series and say which ones should they think about taking into the clinic first out of that series and what would the control population look like and can they create a log or a dossier that says why did they pick the control? And that even sets up their conversations with the FDA, usually even kind of consultative ones at the beginning.
And then how do you optimize that design, so that you have the minimum number of patients required in order to meet all the requirements, the endpoints, biostatistical strength, etc. That's also an AI optimization. And then where do you want to run it? Can you find the sites that actually have the patients that meet the eligibility criteria, but also have the potential to perform on that particular trial? So, each of these are kind of different parts where you kind of get these conundrums and problems and delays and lack of performance, etc. But all of this can be thought of as a system, and this is where I think our AI tools in particular, but AI more broadly, actually can have some real value.
Brendan Smith:
Yeah, look, I think that's a perfect setup for this entire conversation. There's about a million different directions you can go with technology like that. But I think, even for us, one question we get all the time really relates to the data sets being used to train different AI models within healthcare. Where does the data come from? How can you account for differences in quality across data sets? So, maybe more relevant for you or more specific to you I should maybe say, is how do you find the right real world data and ensure that it aligns with the goal for your models or even a given drug given vertical moving forward?
Jeff Elton:
If you were looking through anything that we've made as any kind of press releases or statements, you would see that there's partnering with companies like Caris Life Sciences or NeoGenomics or Guardant, et cetera. And you would find that we acquired a unit of the American Society of Clinical Oncology in December of 23. It was called their CancerLinQ. And so, all of these were initiatives to basically say, can we build a data set that is highly representative of particularly the U.S. population, but highly representative with all racial, ethnic and economic subpopulations that we're comprised of? Because oftentimes what historically people do, they get data set from an academic medical center. So, you kind of know that metro region and that particular group and oftentimes they tend to be a little bit better off than other ones, et cetera for doing that. And it's not unusual. Most AI was usually trained within an individual institution or it's a publicly available data set that people worked on, which is relatively small and now actually relatively old.
So, in part, we bring data in every day. We're bringing in from hundreds of locations kind of around the country. We have to bring in data that isn't necessarily captured in electronic medical record environment. And that's why working with Caris Life Sciences and some of the labs, you're bringing data from their laboratory information systems that sometimes was not necessary just to the interpretation for a standard of care. And you're kind of doing this at a scale then kind with this idea that it is truly representative.
But maybe even more important is the programs that most pharma and biopharma sponsors are working on, they tend to be very narrow in kind of rear population, because they're really trying to kind of prove that it really works in the narrow as possible, because that's easier for an agency to make an interpretation and a decision around. So, that means the funnel, which sometimes is called this funnel where we start with more than 8 million patients, may come down to just a few hundred by the time you get into those narrow characteristics.
So, you actually have to think about that when you're kind of entering this kind of particular domain, that you actually have to have that potential to do that or else you're kind of likely to use a skewed data set or inadequate data set for your training. And then obviously decisions made later on.
Brendan Smith:
And I mean, this leads right into this whole conversation about, not even just which general patients to choose, but therapeutic spaces, indications, which subpopulations of the subpopulations are most right, not just for the treatment but also for the technology itself. And I know, in particular, you all have a few particular areas of focus, obviously you have to start somewhere and then the wide wide world of opportunities that AI is kind of lending to us. But I guess, why oncology? Right? I think there's, again, people often wonder there's so many different diseases out there, so many different drugs and development for each of them. What is it about this area in particular that lends itself to some of these data-driven approaches to trial design and patient recruitment?
Jeff Elton:
There's the emotional, ethical normative side, which is, these are devastating diseases. They destroy families, it destroys lives, it can enter at any age kind of indiscriminately. And I could put economics on it and say it kind of lowers collective productivity by taking people out during prime years of life, et cetera. And it is one of the biggest causes of death, particularly people in kind of working years and things of that nature.
But I think there's another feature which is going back decades, we decided to kind of declare war in oncology. It was actually called the war on oncology. And that got things like NCI funded and a variety of other areas. And you actually started building understanding that these were genetic mutations that were driving a tumor to grow, and all of a sudden we're now getting into the genetics of humans and actually the genetics even of the mutated cells and different organ systems and we understand where it's going.
And so before, but with all of that research and all that area, we had more knowledge of the disease biology and we had more data that was molecular, around that disease biology. It actually had greater wealth of information than any other disease, including cardiovascular, metabolic, or kind of things of that particular nature. So, particularly if you're an AI company and if you actually want to move into something where you're really trying to model a disease state, a patient population intervention and predict which treatment they're going to come to, et cetera. But also the cycles are short, because oftentimes when somebody gets cancer, it may be weeks or months before their life may not be there. And so, if you have something that requires 20 or 30 years to see the outcome, like in cardiovascular disease or metabolic, it also takes you forever to understand whether what you did actually even yielded that outcome.
So, there were all sorts of reasons to say, I can have an impact, I have data, but I'm also going to know whether it worked within a very finite timeframe. Now, you may know, cancer is a disease of different organ systems and has predecessor states. So, we know that somebody who had hepatic cancer may have had liver fibrosis before that, may have had HCV exposure before that. So, because of that, we know a lot about other diseases and we know a lot about the immunological status of patients, so we can start moving into immunology. So, as a place for us to start, we actually believe it was the right place to start a company like ours.
Brendan Smith:
Yeah, I mean I think look, again, you have to start somewhere. And I think when you have this technological advancement that's just rapidly innovating almost on a daily basis at this point, again, it's almost incumbent upon people with the levers of power in this respect to really kind of run with that.
Jeff Elton:
But you also know too, if you look at the pipeline of pharma, 40% of the programs they work on are in oncology. Now partially, they have more confidence in what they're working on, but the next biggest disease category is probably 10%. So, if you're also just sitting here and saying that you're working with biopharma and pharma, as part of that, this is also the area. If you look at breakthrough designation programs by the FDA, 90% of them are in oncology. So there's all sorts of reasons here that it made sense for us to move into this area.
Brendan Smith:
And I mean, fair enough. We also both know that you guys aren't exactly sitting on your hands, right? You're not limiting yourself to oncology. So, I guess when you think about the scope of the technology beyond that, how might some of these oncology focused approaches that you're all using really be leveraged over time to expand as intensely into some of these other therapeutic areas? Are there important bottlenecks to doing that, that you think could be addressed with some broader awareness? What are just some of those considerations?
Jeff Elton:
Great question, Brendan. I'll start on one level, which is, there's nothing about the AI architecture which we refer to as CARA. Cara kind of was clinical and research accelerators is where that term kind of came from. There's nothing about that architecture where we have changed how we manage all data to be appropriate for large language models and agentic AI. So, we kind of went from being more biostatistical, are SaaS focused in terms of how we're doing it. Basically we prepare data to actually be used and consumed by large language models and agentic AI, et cetera. Nothing about that is just purely specific to oncology. So we could transplant that kind of paradigm instead of workflows and kind of things over to what we're doing there. Patient matching, the ability that we actually to work within the workflows of a healthcare provider research site and find patients for trial eligibility.
Everything we do there, we could do in any other kind of disease area. The advancing, the modeling of a clinical trial, the optimization of that [inaudible 00:14:10], everything we're doing there, we could do in another part of it. Simulating a program and a population, if we had the underlying data and the other disease, we could actually build the digital twin and the trial simulation tools, and there's nothing about that that we couldn't move into another disease area around doing that. When we designed these, we designed them to kind of knowing that we have to generalize to diseases that we can't quite anticipate yet. So, that was always in the paradigm about how we thought about setting up the company.
Brendan Smith:
And look, I think we've talked very, very specifically about the clinical stage development for some of these, right? But it seems almost from the outside, from a higher level. If something works in the clinic, you find the right patient population, you're able to design the right trial using that patient population and the drug works see positive phase two, phase three data, that maybe the next logical step would be, these methods could be leverageable in a commercial setting, right? So, is that fair to say, I guess when you're thinking about the transition of these approaches from clinical to commercial, are there any additional considerations that you all need to get two to three steps ahead of where you are now or what's kind of the status of that now?
Jeff Elton:
Again, great question. So, we started off more research and you'd argue clinical development and medical and maybe epidemiological safety. But now, with some of our newer solutions we're looking at, say somebody just got approval for a new drug, they actually need to find the patients that meet the eligibility criteria for now approved drug and actually see the paradigm of the standard care shift and change. So, we're moving a whole new class of solutions over into that particular area. Even inside the workflow of healthcare providers now we actually have something, it's called an RxLink, [inaudible 00:15:52] does notification of newly approved drugs even to the treating clinicians, to actually make them aware that there actually is something that may not have been incorporated into guidelines yet, but that actually was approved for the narrow kind of patient that they may be thinking about how to treat that particular individual.
So, we are absolutely moving that over there. In fact, I think we do work in radiological interpretation as well. And so, we're completely changing the paradigm of what those radiological workflows will look like. Just also because there's fewer radiologists, more retiring, not enough graduating. So, we actually have to augment a little bit of the technologies that support their work as well. So yes, very much going in that direction.
Brendan Smith:
Yes, absolutely. Yeah. I mean look, you're touching on a lot of these guidelines too, right? And I mean this is obviously not so subtly beckons back to FDA, right? And FDA has been honestly very vocal in recent months about their intentions to get behind some of these more recent computational modeling technologies and AI in general. We've really seen a pretty concerted push in recent months to embrace some of these approaches, but as they're also pushing for broader integration of these tools across multiple areas of regulatory oversight and review. So, I guess, how has that recent guidance out of FDA shifted your conversations with collaborators or even just your internal approach to next steps for you at the platform?
Jeff Elton:
I think, you're aware, kind of under Marty, he's brought in Jeremy, who's now the first chief AI officer of the FDA. That January, there was a January document that was preliminary guidance for [inaudible 00:17:24] that came out, that was designed to say, you could integrate AI in support of any form of regulatory decisions and a seven step process built into this. Then most recently you had that Elsa was announced at the FDA, feel like it's a child being born, but Elsa was announced at the FDA.
Brendan Smith:
Well, it is.
Jeff Elton:
There you go, probably is. Elsa came a little early, was supposed to be here on the 30th of June, and ended up kind of actually being almost two to three weeks in advance of what Marty had indicated was around there. And so you have AI in their infrastructure and then you have an invitation to build AI into the programs of the sponsors and use it and actually maybe even to select the patients that you want to bring into the clinical trial or maybe even have it be an AI biomarker that could be commercialized with the therapeutic. So, in my mind, there's been a lot in the popular press about a little hand-wringing and things of that nature. And actually maybe that's why you're asking the question. I actually kind of come on the side that I think we're all gathering experience in the application of AI.
I think it's actually a good thing for the agency to have a walled-off secure generative AI review process that could do some pre-processing, look at precedent, look at other programs that may have been in the same area. I think their safety analyses will be a lot faster, a lot easier. And these are expert tools in the hand of experts. And again, the reviewers have not been dismissed from the agency, the data science team have not been dismissed from the agency. And the new tools actually, they have new staff that are helping deploy these tools. But if you're a sponsor and if you know the agency is going to be using generative tools to assess your program, and if I was any sponsor, I'd want to have mirrored tools and capabilities to know what they're going to know and to see the patterns that some of these tools and some of this infrastructure is going to reflect, so that in fact, I'm well set up to be in those particular conversations to kind of doing this.
So, I think right now, in fact this would be a really interesting thing just to even programmatically spend a little bit more time on. There's some people going into it and saying, this is real, it's there and it's going to change everything about how I work. And there's others, much like it occurs among different sponsors and pharma, say, I want to see this go a couple rounds, I want to understand how this is going to work itself out. The concern that I think people, or at least the interests people should ask themselves about is, if anything, everything with the current generations of AI seem to go a little bit faster and we built a little bit more capability than we thought that was going to be there, and we get a little bit more informed about where we should not be using it yet and where we should be using it. But those cycles are contracting and they actually coming forward.
So, if anything, I think if people are not building that competence and that capability, they're going to find catching up in the environment actually gets a little harder, not easier at the end of the day.
Brendan Smith:
Yeah, I mean this is a great setup into, some of the, maybe I could say the word disparity, but I would say distance, difference between what's realistically happening on the ground, what's happening with an agency, what's happening from the sponsor side. But obviously we exist within an investment community as well. So, I think then the next natural question is, is there anything beyond some of these steps that you've highlighted just now that, particularly the investment community, you might think maybe under appreciates or misunderstands about realistically just what's already happening and what you've already been able to accomplish at ConcertAI, kind of really pulling back the curtain between some of this hand wringing as you kind of referenced it earlier.
Jeff Elton:
So, let me kind of subdivide this between... And Brendan, you cover pretty broad [inaudible 00:21:09] actually in your own work and I want to give you a little bit, you're one of the individuals I read to figure out what's going on.
Brendan Smith:
Thank you, that's great.
Jeff Elton:
So, I thank you for doing that.
Brendan Smith:
Glad somebody does.
Jeff Elton:
I've read it all.
Brendan Smith:
Yeah.
Jeff Elton:
So, I think let's take Therapeutics Diagnostics and some of the other tool companies that are there, because you've kind of covered this. If I'm a therapeutics company, I almost could start thinking about my program is clearly a biological agent. If it's an antibody or it's a low molecular weight chemical agent, and it kind of comes with a whole bunch of reasons why it wasn't designed the way it is, but I'm going to have a layer that's going to understand disease biology, I'm going to have a layer about why I selected the patient population. I'm going to get a lot more information.
So, the more people start understanding that there's going to be models and capabilities, they're going to be as much a part of that program as the biological or chemical entity that establishes the drug itself, that becomes part of things, the better off they're going to be. And in fact, if I had to put a hypothesis out there about where this agency, because they've committed that they're going to have the fastest turnaround time on review decisions of anybody out there and they're trying to [inaudible 00:22:21] that place, what that also means is, because they're using generative AI, surveillance of a program after that decision's been made or approval, et cetera will continue on. And so in a way, all decisions are almost like preliminary. All decisions are almost, except for the data that we will acquire in the future, and I reserve the right to come back and actually have conversations in the future.
So, that's not a bad thing, that's just the thing, that's going to be the structure. If I'm a diagnostics company, and you've covered this in some of your own work, some of them now are AI companies in addition to actually being... So, diagnostic is acquisition of information that gives you an insight into a health status and what drug maybe they're going to be a responder to. So, the notion that some people call it signatures, some people call AI models. I think diagnostics is going to have an AI layer kind of, you're going to have AI middleware and a decision layer, and that's going to be where that industry goes. And it's part of where we've wanted diagnostics go for a really long time and it's getting there. It actually is going to start moving a little bit faster to that. And I think tools companies that produce huge amounts of data, will also find that they're going to end up having an AI middleware that kind of starts to be in between them and then the program that they're working with an individual manufacturer or sponsor around, as we're kind of doing that.
Brendan Smith:
Yeah, look, and I think this entire conversation, to your point just now really is spanning all the different, I mean not entirely encompassing all of them, but really getting a strong sense of the different use cases for some of the technology rights. By no means siloed into one area versus another. Some places might have a little bit more information available, better modeling, but it's almost feeling like an inevitable push forward and kind of the progression of how this is going. So, I think, in that respect, I know we've covered a ton of great ground in many respects today, and it's a conversation I'm sure we'll continue to have for the foreseeable future. But before I let you go, one thing I do like to ask all of our guests is, if everything we've discussed today goes over someone's head, but they've made it with us this far, what is really one point you would want everyone listening in, irrespective of their own expertise, to just really take away from the conversation?
Jeff Elton:
Yeah. So first, and maybe I'll have it be from both of us, thank you for listening all the way through to get to this particular point in that.
Brendan Smith:
That's a good one.
Jeff Elton:
I think the second part is that, in healthcare and in life sciences entities, this concept and this phrase of AI for good, actually my very strong belief is that, people inside pharmaceutical and medical device companies are making the drugs and doing the work because they want to actually better the human state and they want to do things to truly either stop a disease or better diagnose or get something to it. And so, AI is an accelerator for the things we've been trying to do and can actually act as an assurance. And it is AI for good.
It is not, I actually think, and the only way it'll get refined is refined in use with transparency, with publication activity, which as an industry we're good at. Think about your own analyst reports [inaudible 00:25:26], and you can track publication activities and other. So, in a way, this is not scary things. This is actually taking away that humans have thought. I mean, an LLM is a series of neural networks which are patterned on how we think as human beings in kind of doing this. So, these are accelerators on the way we have considered and thought and did things in the past, but really to help people's lives actually. That's actually the key takeaway that I'd like people to have.
Brendan Smith:
Fantastic. And I think, with that, I do want to thank you also for hopping on and talking us through this entire process and just all the evolving use cases for AI. I'm sure we'll have plenty more to discuss over the weeks and months ahead, as more of this work continues to bear fruit. So, big thank you to Jeff for joining us today.
Jeff Elton:
Thank you very much, Brendan.
Speaker 1:
Thanks for joining us. Stay tuned for the next episode of TD Cowen Insights.