S1, E34 - Matt Truppo, PhD, Part 2: AI-Driven Drug Development at Sanofi: Clinical Trials, Regulatory, and Personal AI

Welcome to Practical AI and Healthcare, the podcast that cuts through the noise to spotlight real-world solutions delivering real-world value. From patient care to clinical research, from life sciences to patient engagement, we focus on what's truly moving the needle in healthcare. No hype, no theory, just practical insights where AI is making a true impact. Look them aboard and let's get to it. As many of our listeners know, Leon and I work very closely with the DCI Network and Division of Clinical Informatics at Bethesional Deakin's Medical Center in Boston. This June, the network is hosting patient-powered digital health 2026. The conference will bring together patients, innovators, industry leaders, healthcare providers and policymakers to shape the next generation of real-world patient-centered solutions. The meeting will run from June 22nd to 24th in Boston at Harvard Medical School. We've arranged for our listeners to get a discount on registration to the meeting. If you register between now and May 15th and use promo code Practical AI June, no spaces, you'll receive 30% off your registration fee. You can learn more at DCINetwork.org/Petians2026. In addition, we're always looking for sponsors. If you are your company or interested in becoming a sponsor, please reach out to Admin at DCINetwork.org. See you in Boston. Hello and welcome to this week's edition of Practical AI and Healthcare. This is going to be a part two of our conversation with Matt Truppo. Matt, as you may recall, is the global head of research platforms and computational R&D at Santa Fe, one of the larger pharmaceutical companies in the world. If you haven't heard part one, here's a short version. Matt walked us through how Santa Fe is using AI in the earliest stages of drug discovery. How they're finding not-full drug targets, designing molecules, engineering proteins. Some of the results are pretty striking. They were able to identify 10 new targets in 12 months and 30 million target pairs that were screened in days with protein engineering time cut in basically in half. These are some really impressive statistics, Matt. We want to hear more and we're going to flip things over in a sec. So where we left off was right at the edge of the clinic and at the edge of our seats. So Matt described the holy grail of using preclinical data to predict clinical trial outcomes. That's where we pick up today. So clinical trials are regulatory strategy and then we're going to get to a story about Matt turning himself into AI. So lots of fun stuff coming up. So Matt, welcome to the podcast today. Thank you very much, Steve. Thanks, Leon. It's good to be back with you. So Matt, last time you described the holy grail, the digital patient twins that integrate all the preclinical data to predict clinical trial outcomes. How close are you and where does this stand today? Yeah, so you talk about as the holy grail and it's true. That would be sort of the ultimate use case of leveraging AI and advanced digital tools to accelerate what we do. And I'll put it in context and then I'll kind of talk a little bit about how close we are, what results we've had, where we still have a lot to go to get to that holy grail. So what's the problem statement as you go into the clinic, right? So about 90% of all drugs fail during clinical development, as you know, and it happens to be the longest and most expensive part of our R&D process. So not really where you want a huge failure rate, but that happens to be where it is. And most of that failure is due to lack of efficacy. So if we can get even a little bit better at predicting a drug's efficacy, at predicting a patient response, and we could reduce attrition by let's say 10%, that would effectively be a doubling of our industry's productivity. So that's the magnitude of what we're trying to achieve. And we're definitely not there yet. But what I can do is show you some examples of approaches we're taking that's trying to help us drive progress in that direction. So I'll give you a couple examples of where we've actually had concrete success and the approaches that we've taken. So in digital patient twins, one of the methodologies that we've leveraged, well, QSP or quantitative systems pharmacology. And these are models that connect known pathophysiology and a drug's mechanism of action to observe biomarkers and endpoints. So this is powered by differential equations, ordinary differential equations. And it's quite powerful. Because if you can build these, that's called the replicas, or digital incarnations of individual patients, and then resulting patient populations, you can effectively test a drug's response before you ever put it in that patient to see, is it going to be differentiated versus other therapies? Is it going to be something that we should continue to pursue in that very expensive and time-consuming process of clinical development? And we have a couple examples of where we've been able to deploy this effectively. The first one I'll talk about is our most promising drug, which is currently approved, which is dupexant. And this is an IL-4 receptor antagonist antibody. There's an indication for chronic rhinocytocitis with nasal polyps. And in that indication, we were able to build population pharmacokinetic models and pharmacodynamic models using QSP to simulate an adolescent response based on our adult clinical data. What that approach enabled us to do was extrapolate the efficacy and safety profiles from adult to adolescent populations without the need to conduct a dedicated pediatric trial. And the reason is because these are explainable mechanistic simulations. So that data bolstered the support that we got for our regulatory approval to expand into an adolescent population. Effectively, it reduced our cost, shortened our development times, and it minimized the patient burden, expanding access to new patient population that was in need. So this is something real that was applied to our largest product at the moment. Yes, a math, really exciting story. I want to probe on two dimensions of what you just described. You mentioned that QSP is usually done with ordinary differential equations, right? The old-fashioned math. Is there an AI angle? Are you using newer neural networks based approaches to enhance us somehow? And if so, can you talk a little bit about the architecture? Yes. So what I can talk a little bit about is where we'd like to push in future directions. And some of it, we're not quite there yet. Some of it, we're starting to expand. So there are two approaches that we'd like to apply more advanced digital tools like AI to. And one of those is expanding the ability to do QSP-like simulations much faster, because they're very compute-intensive. And it actually takes quite a bit, small army of people, to build those models for each individual disease state. And they are specific to each disease state, which means that you have to build them out for every disease and every indication, and then apply them to the digital patient twins that you want to build. So one approach is to say, can we shortcut some of that by having AI learn from these QSP models to more effectively and rapidly build what are effectively mechanistic models? And we're doing this in the QSP space. We're also doing it in, I didn't mention it in the first episode, but for quantum simulations as well, taking classical quantum simulations that are very compute-intime-intensive. Can we work with partners to say, let's build AI models that can shortcut that in a much more compute-efficient way to expand that to earlier in the pipeline? And so I'll talk a little bit more about where we'd like to push AI also. But I will just say there's one other example I wanted to highlight where this type of simulation has helped us out significantly. We're a company that works in rare diseases a lot. And so one of the, and I'm really proud of the fact that we work in a space where we have a long-standing commitment, life-changing, life-saving therapies, an area where the patient population is much smaller. The challenge is that drug development in rare diseases is more challenging because there's limited number of patients, they're geographically dispersed, so it's hard to find them and recruit for clinical trials. And there's a lot of patient phenotypic heterogeneity as well. So that can lead to variable disease progression. So this is where also things like AI looking at real-world data, real-world evidence combining that together with QSP can create really valuable digital patient twins. And the one example I'll give in that space is we had built for a product which is again now on the market, Alipadase Alpha. This is Zempozyme. It's for a really, really devastating disease, acid, sphingo-myelinase deficiency. It's a fatal disease. And it's a disease that we were able to build QSP models and again accelerate our rare pediatric drug development. A phenomenal example of where QSP modeling and digital patient twins were applied. We subsequently were invited to present that at a workshop to start to create a road map to how do we leverage these types of models in drug development, leveraging digital patient twins for approvals. So it was an approach that ultimately allowed a treatment for a fatal disease to reach pediatric patients faster with fewer children exposed to an experimental condition, experimental therapy. So these approaches are really powerful. Now you asked about where do they fall down, right? Where are we not there yet? Well, I was going to ask you, I was going to ask you that back straight to it. Absolutely. So, and this is where we're hoping, but we're not there yet. AI tools can start to bring together more multimodal data like real-world evidence and other things. So I'll talk a little bit about that. So the benefits of the models I just described is they're not a black box. They combine pre-clinical data like in vitro assays and vivo studies with prior knowledge of the pathway of that disease, the biology of that disease, and clinical data sets, either from competitors that are published or ourselves. The drawback is you need a good map of that disease's pathophysiology to build those models and you need some human data with respect to the target of interest. So the big gap is how do you build virtual patient engines or digital patient twins for first-in-class targets, and that's where the gap is. And that's a pretty good gap, right? This is really important to get into first-in-class target space where there is no prior human efficacy data at all on the targets of interest. So what we're doing there, and this gets to the second part of where we're trying to apply AI, is, and we actually, in fact, have an institute, the Biomedex Institute in Heidelberg, Germany. We have a partnership with them to build out some of these capabilities. And it's a biology first modeling of causal disease mechanisms using AI. And so this is based on in vitro data, based on in vivo pre-clinical data, also based on real-world patient data. And we want to combine these data sets along with new systems that are coming on board, like organoids that maybe and could represent an orthogonal way to get interesting data on actual human cells up to the tissue and cellular and organoid level. And the idea is that you could bring that data together to start predicting response to a new drug, a new therapy, with first-in-class medicines. So that's something that is not built yet. It's something that we are working on with external partners and experts in AI experts in these multimodal data sets. But that's the next generation of where we want to push with this technology. So really exciting progress, Matt. And thank you for sharing an actual specific example of an achievement. And you know, Steven, I have both worked in rare disease registries, and I've worked in a few that were ultra rare like Spino muscular atrophy SNA, which I think you guys had a early drug for. There are 400 patients in the US. I mean, getting a clinical trial run on babies, where there are 400 of them is not easy. And anything that could be done to improve that is really, really important. It's impressive that you guys were able to actually extend the indication from pediatric to teens with just the digital tools. Now, one thing we know about drug development, that you know, of the many things that makes it difficult is like 30% of drugs fail in clinic due to unexpected talks. How good are the twin predictions in getting around the toxicology of predicting the increased talks? Where does it and again, where does it break down? Yeah, so there's a couple of approaches that we're taking to the talks piece. The first is looking at the properties of the molecule itself in preclinical sprays prior to even building digital patient twins. And there we can start to narrow down and identify what molecules would be more susceptible to talks and then evolve those molecules away from that. So these are things like herg liabilities, where we are moving to a system where we've actually got a series of co-crystall structures of small molecules with the herg receptor to look at QT prolongation, things like that, cardiac toxicity, and now simultaneously in all of those small molecule projects, we're actually looking at simultaneously fine-tuning the molecule for the target of interest while detuning it for any herg liabilities. So this is something you can start to do in real time in the discovery space. And then there are other like dilly for drug and deuce liver injury. You can start to look at molecular features that coincide with that as well. So that's in the sort of before you build the digital toon space and it's part of it is bread and butter but part of it like the detuning for herg liability. There aren't many really good data sets for the herg receptor in terms of high resolution crystal structure. So to be able to have that and we have cryoam structures on this, that's where we're starting to build AI-enabled tools based on these structural data sets which we didn't have before. So then once you get to the digital patient twin space, now you're looking mostly at how do you differentiate on target toxicity. So we're trying to dial out off target toxicity in the prior examples of where you're designing the molecule. But now you're saying, well on target, how much can you hit that target and have a beneficial effect versus start to have a negative effect. So that's where the differentiation for basically where we're looking at these tools. In terms of the practical, how much of an impact it will have on the tox piece, it remains to be seen. I think that there's good indications that we're headed in the right direction to detune some of these things. There's good indications that we're starting to be able to predict things like. So another good example actually was with a drug that's in the clinic now in second meag. This is an anti-t slip I-23 multi-specific nanobody. This drug candidate in the context of asthma, we actually looked at predicting treatment effect on biomarkers like eosinophils, IgE, like fractional exhale nitric oxide, which is a good measure of airway inflammation. And so both of those measures, both the endpoints as well as the biomarkers relate to efficacy of the drug, but also to potential toxicity. And so by looking at those, we're able to predict successfully the phase one B results from the digital patient twins we built. And that allowed us to actually go directly to phase 2B, skipping phase 2A. So we didn't need to do a dose finding human cohort because we predicted the dose that we needed to move to phase 2B. Now time will tell as these trials go on what the readouts are, we don't know yet. But we do know that we were able to go in with a dose that was predicted rather than spending the time and resources and patient time going through that phase 2A study. So these are ways that we can start to address these issues, but they're not wholly solved yet for sure. So, Matt, as we've been talking here, which one I just framed things for our audience. Last episode was really about the R in the R&D space. This episode is really about the D, the development. And I'd like us to pivot a bit over to the D space. That's an area I've spent a considerable amount of time in my career. And let's pivot around to some of the places where rubber hits the road in terms of the actual trial execution. You know, in my last role, I was at a large pharma working in clinical trial feasibility. I ran the analytics group for a large pharma. And we were looking at these issues around trial forecasting and better site selection. And to be honest, at least on my watch, all this stuff was just kicking off. We knew it was something that we could do. We wanted to do it. We hadn't really hit our stride yet. Where are you guys at this point with these really, you know, rubber hits the road, kinds of activities that, you know, if you can improve them, can really have a market effect on trial execution and getting a drug out to a regulatory submission. Yeah, yeah, no, it's it's a great question. So what we've done in the clinical trial space, it's actually, it's actually pretty remarkable the last 18 months in particular. So I'll describe to you what we call our clinical control tower. And what this is is a data hub and dashboard. So both the visualization as well as the single source of truth for where the data is for the entire clinical development space. And I'll kind of describe what we use it for and what it's what it's been able to achieve. So this solution, this clinical control tower has replaced over 15 legacy applications where folks had to go to many different places for a variety of things, which I'll tell you about. And as of the end of last year, it's been rolled out to over 4,000 users. So this is now its scale. This is deployed at scale in the company. It helps us do a couple of things. You mentioned site selection. It helps us prioritize sites based on the institution's experience and their performance metrics for a selective therapeutic indication. So things like what does the legacy startup time been, the recruiting rates, the screening failures, etc. So when things when folks are thinking about in a specific TA and a specific disease recruitment, where should we run these studies, all of this information can now be placed in one dashboard in front of them and they can drill down on a global, original and site-based level for all of those metrics. It also, once a trial has started, it tracks patient recruitment and enrollment globally and monitors site activity at each of those sites. So the continuous monitoring of site activity and investigator performance and flagging risks like is a site having a high drop outbreak is the recruitment trajectory at risk and that then would put the readout at risk, which would then put the filing at risk. These are automated flagging of risks on a site by site level, on a program by program level. So by tracking those trial health indicators, it can help us support decision making because you can imagine and as you know in the industry, you're making decisions between are we going to move forward with Project A or Project B in what time frame based on how we can recruit, how we can, the asset that we have and the medical need. So these all come together in terms of real decisions about what you're going to be progressing in the clinic and beyond. It also combines in that clinical control tower clinical supply forecasting. So we can include things like the shipment specifications of sending drug and formulated material to the site. sites that need it, when they need it, as we look at when individuals and patients are going to be coming on to a drug. The workload forecasting is also present in that clinical supply chain forecast. So this can help with things like reducing discard rates of material that you didn't use, making sure that drug gets to the place it needs to in the right amount of time. And then I mentioned the performance dashboard where you track all of those key metrics simultaneously, and that ultimately leads to what is the cycle time. How fast are you able to execute on these trials successfully? So we're doing quite a bit in that space and rolling it all into one single dashboard, which interesting enough people can actually access those 4,000 folks on their phone, on a mobile app as well. So it's quite a bit has been done in the last 18 months I'd say. Yeah, the thing that I'm going to probe you on this a little bit, because a lot of the things you just said have been done in the absence of AI. How is AI actually doing all this? I mean, one of the predicating things we were working on was getting all our data harmonized in a way to do what you're describing. But at the time, at least, they weren't predicated on AI tools as you're describing it. So what's the AI kicker that you're getting from doing everything you just described? Because those things that you'll agree with me have been done for maybe 10 years. What's the new thing here? Yeah, so some of it is certainly legacy approaches, non-AI approaches that are now harmonized into single system with single dashboard views. That is the big part of it is that data piece and that data analysis piece. And then some of it, like the risk flagging, is AI enabled where you're looking at multimodal factors. So risk flagging could be something that is related to geopolitical risk with chipping lanes. And we've seen that recently. Yeah, really? It could be something that's related to making sure that drug is arriving on time. It could be something related to the trajectory of recruitment and then how that cascades into when a readout might potentially be. Because we all know that that is a guess. You're guessing when a drug, when a trial will read out. But all of those guesses feed forward into your portfolio management. So we didn't talk about that. But that's enabled by AI as well to say from a portfolio management perspective, beyond the classical sorts of analyses and sensitivities, how can we take all of this into account to have risk associated with our programs reading out, not only technical risk, but also the justical risk, factoring to our decisions. So the AI layer is really about bringing all of those together so that we can start to look at it on a program and portfolio wide level and make better decisions in that way. That's where the AI piece is coming in. Really nice to hear. As I said, I've had a lot of experience in this space, but it was earlier on in the lifespan of where we are with AI. So we didn't quite get to the wins that you're describing. And it's amazing that, hey, you're getting them in B that you're able to talk about them in a public way like this. A lot of companies would consider this to be proprietary. So thank you for that. Let's shift a little bit because the use of biomarkers is something that is really important. It helps us figure out what drugs are going to be used in what disease states are being able to measure the outcomes of a disease therapy. How, in our pre session, we talked about biomarkers and how they're working and how AI is helping. Can you expand on that a little bit for the audience? Yeah, yeah, sure. So I think the use of biomarkers is really critical, particularly when you think about you're trying to identify, I mean, in a simplest form, what is a biomarker? It's some test, hopefully a simple one, but not always, that you can relate directly to a patient outcome. You know, you're looking for something that relates directly to patient outcomes so that you know if the patient on therapy is headed in the right direction is having a positive response. And it's something that you can practically test so that in the clinic, you can really measure and see which direction that patient is heading. So it's, I think it's super critical to how we execute on our programs. It's something that you need to demonstrate both that the biomarker is relevant to the disease and the disease progression. And you also need to be able to identify biomarkers ideally that I mentioned are easy to assess. So that the trials can be run in a more streamlined way so that it can be easier for patients that are in those trials in terms of patient convenience that to participate. So that's all quite critical. And I think that what's important is how do we identify novel biomarkers that are relevant to those diseases? And this is still early days. I mentioned, you know, I mentioned in the case of, in second, maybe a phenobion marker. That's a very, I'll call it a, a poignant and relevant biomarker which tracks really well with patient response and airway information in the case of asthma. And so that was something that you could look at drug and then assess the, and predict the impact of that biomarker. And our digital patient twin models lined up extremely well with the patients we were seeing in study. So that was something that gave confidence that this is a good biomarker, one that we can model and then one that we can follow patient progression. As we try to identify new biomarkers in different spaces, this is where the use of tools like AI to look at rear-wheel data, rear-wheeled evidence, and really see are there biomarkers that we're not using today or that the fields are not aware of that are relevant because they are popping up as patients progress through their disease and through various stages of their disease. And they have correlated biomarkers that we didn't find before. And these could be things the, the place where, you know, you most commonly look as in blood because it's blood plasma, easy to get and easy to search through. But it can also be in other, in other areas as well. So this is something, again, it's a, it's a new space, it's an emerging space. You need to crunch through a lot of data to get to the signal through the noise sometimes and this is something that AI can be really good at. Yeah, absolutely. Canon, in fact, you mentioned signal through the noise and I'll just plug our conference because Leon and I are on a committee, well, where were the planning committee for this conference at Harvard in the fall called, in fact, signal through the noise. I don't know if you picked it up from the last episode or not, but this is the nature of what we're doing up at this program at Harvard. It's pivot to the regulatory environment because, you know, everything you're describing and doing eventually is going to hit the FDA in one way or another and the FDA or the MA, they're going to have to be happy with everything that's going on here. They're going to have to approve it and be content that the findings that AI is helping to bring to bear are trusted, are accurate, are precise. And where are you guys now with those challenges? I know FDA is trying to move, but it's a government agency and they don't move at the pace of AI. What are you seeing at the moment? And is it curtailing you or are you able to move it forward faster? Yeah, there is definitely a willingness at the agencies and it's really borne by the same goal, which is to accelerate the progress of novel patients to medicines and to do it in a way that is less costly, right? So there's a common goal that the agencies have and that the industry has. And I would say that there is a willingness to say, and you've seen this in some of the new releases of guidelines with respect to modeling and simulation, where there's a willingness to accept modeling and simulation provided, you can explain how you did it, right? That's important. And that the outcome is reflective of what you've modeled and simulated. And so I think that makes perfect logical sense. So the approach that we're taking at the moment I mentioned QSP models, right? These are models that are more deterministic. They are mechanistically accurate. And so those are models that are more straightforward to explain. We haven't gotten to the stage yet of a model that is sufficiently black box that we have to wonder, okay, how are we going to explain this? That's going to be something that we're going to have to work on together with agencies because ultimately you need to be able to stand by the findings of your modeling and simulation. You need to be able to demonstrate that it is accurate, that it is safe. And you need to be able to have some explanation for why it works, right? And the methods you use. So I think the important thing is what is the data going into that model that needs to be transparent? What type of model are you using? That needs to be transparent. And then does that model recapitulate data that you're actually seeing? And I think with those three elements, that's how we're driving forward. And ultimately it's worth saying that every decision that we make is in the hands of a human being, every decision we make about putting a patient on drug, about what the dose is, about how we interpret the data that has to firmly remain in the hands of expert human beings. These are all tools to help us make better decisions, but they do not replace the expert in the room. So Matt, you've described some really interesting patterns so far that I want to highlight for our audience. One is you've described a pattern where AI is using the accelerant to create, to more quickly create deterministic analytics, right? Like with the QPS. So you can write AI can write systems, differential equations, better faster and discover, you know, more comprehensive ones than human teams, for example. Then there are other kinds of systems where you're using narrow networks in a more direct way, while we don't want to call them a black box, right? They are less transparent, less explainable, less justifiable. Now, when we talk to you about this in part one, you made a really important point, which is what if a model is 99.9% right, but is unexplainable, that may be fine in low risk preclinical work, right? You'll be like, look, I, you know, how much toilet paper did you order? you know, as long as the answer is right, I don't know if it's necessary for us to nail down exactly the deterministic principles, but what you did it. But clinical trials aren't low risk. So how does the explainability bar change depending on how close you are to clinical use and to human safety? Yeah, the explainability bar change is quite a bit. You need to generate evidence in a sufficient explainable way in order to be confident that the next step you're taking makes sense for patients, makes sense for both safety and the potential efficacy that you want to see. So the explainability bar does change quite a bit. And I think that's something that's really important that you know, we mentioned earlier in the first episode that the comment that you said, if it's 99.9% right, you know, in a space that ordered in toilet paper. I mean, I would even say if it's 99.9% right in predicting a synthetic root to a small molecule, that's probably fine too, you know, because you're always going to make sure that at the end of that synthesis, you then use definitive deterministic analytical methods to say, this is the purity, these are the impurities present, this is the levels. So you validate all of that. And that's what really goes into the safety, the efficacy, and the formulation of that drug. So I think that what it comes down to is you can take non-deterministic, you can take models that are a bit more back box to help you answer a question. But the ultimate how you answer that question in a patient, how you answer that question to a regulatory agency, it has to be grounded in evidence that you actually provided that you actually demonstrate. And that's that's kind of think I, you know, we're, we're things stand today. A good example is actually when you talk about, and, and we might, we might go here next, actually in the conversation, but when you get now non-clinical into regulatory document generation, right? This is a place where a lot of people are playing. And this is Gen AI. I mean, this is a, a black box, if you will, large language model, generating documents that you're handing to agencies. So how do you deal with that? So, so I'm happy to get into some of that as well, because that is a, that is a black box environment, but, you know, how do you deal with that? Because it's very powerful for writing documents. So we could talk about that if you'd like. Yeah, let's get to it. And I want to say that the concepts you're describing, I think, out there in the zikai style, you know, one of the things I'm working on is the drug information associations, AI consortium. And when we started bringing people together to talk about use cases and risk, it became very clear that risk is very use case bound. And we're, you know, we're talking about risk in context governance, right? Not just risk in general. I think that's better aligned with the FDA than the EMA, but we can leave that for a, for a fight when we have regulators in a conversation because they got to defend themselves. So, but bring us, so you talked about the handoff problem before, right? And great models with manual handoffs to the next team. And I think that happens a lot in clinical ops and in documentation in clinical ops. How many handoffs exist? And is it genticae realistic given the regulatory constraints for those use cases? Yeah, so I think, I think that there's there's always a lot of handoffs in our industry, right? There's a, when you think of the sheer number of experts in each discipline that need to touch the process to go from all the way from target idea where we started in the last episode, all the way through to now where we are in terms of talking to regulators about filing. And that will always be necessary from a human perspective to get the human decisions in the loop. However, what we find and I think this is a common problem is that when you transition the data used to make those decisions from one system to another and you do that many, many times, dozens of times, that's where you have a lot of white space that you could eliminate if it was done in an automated fashion. And so one aspect of that is as you're coming out of the clinic, you have data that you need to compile for clinical study reports. You have data needed to compile for the comment technical document, which all goes into these documents that are thousands of pages long, but the data comes from experts from medical, from the clinicians, it comes from the preclinical space and preclinical models, it comes from the translational medicine teams, it comes from CMC who makes the actual drug substance and drug product. So all of that information has to come in together to feed one document that goes to a regulator to eventually approve. So that's something that takes an extraordinary long amount of time and it's not research, right? This is after all the research is done, after the development is done, it's just to get the document together. And to give the audience who may not be aware of some timelines for this, a clinical study report of which we write hundreds a year. And this is a small section of eventually the comment technical document, right? A clinical study report could take 17 weeks to write, single one. We've reduced that time significantly using Gen AI and this is a space where many different farm companies are working in. And we've reduced it by about 35% so far with a goal of about 70% reduction in time this year in 2026. And what we're finding is that and the way we've done that is classical large language models like off the shelf ones that have been built and then fine tuning them on synoffi data, fine tuning them on our specific data or specific way of writing or specific trials that we're running. And the goal there is again to reduce the amount of time that someone needs to spend in those first drafts. And what we're finding is about 80% of the first drafts are high enough quality already for submission for those for those sections. It's not applied to every section yet. There's four or five sections that we can write in this way that have been scaled across the team. But we're kind of taking one bite at a time of that document until we can get our entire way through it. And as we said before, Gen AI is a bit of a black box in terms of elements, how they spit things out. So each of these have to be reviewed by experts for the for compliance, for completeness of data. But I actually think that where we will land is far more complete, far more quality check documents than we ever have because again, it's thousands of pages. So to have that first draft done in a highly automated way and be able to focus your experts on ensuring that everything in there is not only accurate, but is well described and is complete is really the place to put those experts attention. Yeah, so Matt, if I am a health system that's partnering with you on a trial, what changes for me? Does the AI driven trials as I mean site experiences something different? Does the site burden change for better or for worse? So I think that there are a couple things that I think certainly change for the better. One is the way that we integrate and collect data is smoother because we're able to ingest information from those sites faster and I think in a more in a more direct way. The other thing that changes is support of those sites from our various clinical medical teams, we can predict better now which sites are going to need support and when because we can look at the trajectories in real time of patient enrollment and then that can feed forward to say, well, then, you know, when are people going to get on study drug? When are they going to need first dose, second dose, et cetera? And so by doing that, we've actually been able to take our teams that interface with the clinicians and we've been able to predict better when they will need support, when they will need visits and that way they've gotten better support with the limited staff that we have. So in those ways, it actually helps you provide support to the site and ingest information to the site in a more profound and streamlined way. So, you know, the decrease in site burden is profound. I mean, one of the things that we were doing at my last firm was trying to figure out ways wherever we could to decrease site burden because it was, you know, if the groups don't want to work with you, you know, if your protocols are too complicated or interacting with you is too hard, it makes it really tough to get them to help with patient recruitment and running patients through your trials. So kudos to you because that's an area that was really challenging for us. I want to pivot his back to a different part of the regulatory equation, which is that around the mechanics of actually filing, making a filing. One of the areas that we were doing prior to my last role was, and we had some really good success here, was the actual mechanics of the filing process. We could use AI to help generate the package and ensure monitoring what was missing or what needed to be chased down and all this type of thing. And I think you're doing something similar. And this is among the cases that are considered, quote unquote, mundane, but they work really well. And we've been finding anecdotally and some of the folks we've spoken to that these mundane use cases really do have some significant wins. So could you comment on that a bit? Yeah. And you're right. I think there's an area where everybody is pushing because it kind of makes sense when you think of the really successful LLMs that are out there, whatever they've been trained on, they've been trained on a lot of text, right? What are they good at generating a lot of text? And they do a lot more than that. But those are things that they're really good at. So we, like others, have looked at CSR writing, like I said, clinical study reports, the broader comment technical document they were focusing on. But also we've been starting to look at how can me more rapidly respond to regulators questions in a way where AI looks at the question, is able to then collect the relevant information and data to draft a response? Because this is a way that we can be better partners with regulators for faster responses. and more complete responses. So it's funny because you can imagine a world where eventually you'll have regulators, AI, is talking to farmer companies, AI's and they're asking questions and they're responding and then you get everybody in the room at some point. And obviously we're not there yet, but we are having some success with being able to draft responses to questions in a much more fast and complete way using AI to help with that response. And I hope you're at the answer to this is of course, but when you're having these automated responses, there's human eyes on these before they hit the agency correct? Oh yes, yes. So when I say automated, it does not respond in an automated way. It helps you to start collecting the information and to start drafting that response. But the response again is at least currently the response is always from a human being. It is from an expert who is responsible ultimately for that response. So I think that's that's the important thing. Yeah. And how is the agency responding to this? I mean, given that they have constraints in terms of their bandwidth, if you're able to risk there's been a cadence, right? There's a cadence of push and pull and when they send you a response or they ask you for data, the expectation isn't that they get it tomorrow. The expectation is you've got a period of time. And now you're able to compress that time and all of a sudden when they ask for something that comes in faster, are they responding? Are they able to deal with this increased throughput? So, so we've not seen any issues in terms of dealing with the time response and throughput yet. But I think that that's where it'll be interesting to see do the agencies move toward building tools to sort through the responses in a faster way as well. So, so I think you're going to see on both sides and what's going to be important there is that we're transparent with each other on how we are building these tools on how we're building responses. I think that's what's going to be really an important key factor is so that we both understand what's going on in the back and forth between agency and company. So, so yeah, I think I've seen nothing yet in terms of a slowdown. I think it's it's helpful. But in the future, you could imagine like you said that you could get and generate a larger volume of questions, a larger volume of responses. And so how do you deal with that? And I think these are tools that are well suited to deal with that. So Matt, we spent the last two episodes talking about institutional and use of AI, but you told us you personally have been an AI guinea pig and tell us about the digital version of matrico and these multiple expert agents that are now part of your daily workflow. Sure, sure. So, so, so I'm this started as a bit of a bit of a challenge over last summer and part of it started as a bit of a bit of a bit of a bit of a, I won't say a joke, but a bit of a fun diversion. All the best technology innovation starts as a joke. I'm convinced it's true. It wouldn't be funny if we yeah. So really over the summer, I challenged myself to say, could I cut the amount of time that I spend in my day job in half so that I could refocus those efforts on the stuff we're talking about now on the more advanced digital AI applications on these these these cool new emerging technologies and capabilities, both tinkering, but also really spending my mental effort in that space, a bit more profoundly and a bit more volume and that mental effort space. So I can already tell you I like to set really challenging goals so the spoiler rated alert as I failed. However, I got pretty close in the first iteration and so I was kind of describe, you know, what was done and where where things landed. Basically, the concept was a series of custom agents and the ability to do or support me in various aspects of my day to day job. So I started off and I just started in a Microsoft co pilot architecture because most of my work is in Microsoft ecosystem like outlook and one note in teams and share point, but it's certainly not the only place where you could build this sort of thing. And what I learned in the very beginning was that not surprisingly, how you handle data is key. We talked about data in the first episode. So I spent the first couple weeks actually organizing my digital footprint. So things like my communications by email or by teams, the files that I had in my one drive on my desktop creating automated transcripts of meetings in a way that I hadn't really done before with things like zoom and others. So taking that I thought, oh, this is a good start and I started out to create like one super agent to rule them all right and that failed miserably. The reason was it really wasn't specific enough in any single area to represent what I needed to deliver in that area. So as a generalist, it just wasn't good enough to really meaningfully impact what I do day to day. So I moved to a different approach and you're seeing that this is something that I think the industry in general is moving to especially in research where you have very hyper specific use cases that you need really probably small but efficient models. So I moved from a single model, the digital brain to really a series of specialized agents and an ecosystem of them. And then that I kind of approached it was with the help of AI I was chatting back and forth with various chat bots to say or various LMS to say how could I do this how could I set this up so it's interesting to for me one of the most interesting things was to learn how it thought it would do it itself and how that influenced how I then built it or co built it with these LMS. So I looked down my job description into a real job description not the one that you go out and hire with but one that is what do you do day today right and so I prepped for my day in the morning. So there's a morning agent assistant that helps me prep based on all the things I need to read right I lead governance meetings. So I have a project progression agent that knows about those governance meetings about the data behind them about the teams that are coming what questions they're going to need answers to I'm a people manager so I have a one on one meeting agent that perhaps be for those meetings that has the notes from previous meetings to make me a more effective people manager. I helped part of my role set scientific strategy so there are agents that are technically accurate and technically precise in each subgroup within my department. I spend a lot of time responding emails so there's an email agent that's trained on over 2000 emails I've written over the last few years to respond like me not in an automated way I check it but to at least have a draft and the end result of all of that was a series of agents that saved me 30% of my time and that is based on an AI analysis of my calendar in the six months prior to going live with these agents and the six months after going live with these agents. I think I get to 50% but 30% was was quite profound and in doing that I was able to redeploy that time to the things that matter more and focus my time on things that matter more. So Matt really interesting pattern and I think you're reflecting what the industry is learning which is sort of a god agent or a god skill don't do things that are specific enough and the right approach is to break things up into really discrete chunks and that puts the burden on the designer or at least their AI. So I have to figure out what the right chunking is when you're doing the meta automation really interesting for sharing that one thing I'm sure our audience is curious about is the more high stakes use cases like the governance coexistence right you've got you've got them in a high stakes meeting the new ones decisions what is the AI do in that setting and what can to do. Yeah, so the first thing it can't do or shouldn't do is actually make the decision so that again the decision lies or whoever has the D in that meeting. But what it can do which is I'm finding to be quite remarkable is I will I will chat with those agents during the meeting and if you ask things like if it's trained on sufficient data you can ask things like in early drug design efforts. What are potential on target. Different tissues that could lead to talks for this particular drug and you can get a pretty coherent answer you have to watch out for things like hallucinations and all these things that we have to look out for but you can mitigate those by limiting the space that it has access to reason over so instead of having the entire internet you can limit it to a ground truth set of data that you have for those specific programs and that's the way that that I leverage them. But you can also ask things which I found to be quite remarkable where you could say just out of curiosity what would you decide about pre investing in CMC for this program in comparison to what else we would have to slow down so what is the opportunity cost and instantly you can get a response of these are the programs that are slated for CMC batches that are going to be run in the next three months. You would have to slow one of these down and that would impact this and so you can pull together not just decisions about the science of the molecule but logistical things that influence your your governance meetings as well. And so from that perspective it can be quite profound to pull that information together cross programs which is something you often miss in a project governance meeting is you often miss the cross program impact of a decision you make in that specific government. Yeah, Matt sorry Steve love the discussion and actually if you want to bring Matt AI into this meeting I think our audience would enjoy hearing from him. But you know while you do that I'm just going to make an observation which is a lot of executives talk about a transformation but keep themselves out of the experiment and you you showed real leadership and boldness by putting yourself in first it you know I'm going to it sounds to me it's intentional as part of a pattern. And the fact that you just got back from launching an initiative to station f in Paris which is the world's largest start up in computer with AI challenges for external companies it feels like the same. instinct, which is going outside your walls and put yourself out there. And I suspect that's one of the things that's allowing you and Matt AI to be successful in the domain. So if you want to show, do it just a quick show and tell, I think we got 90 seconds for that. Sure, sure. I can do that. So this, what I'll do is I'll just show you, you know, digital map, this started off as a bit of a joke. But when you think about time creation, how can you be in two places at once, right? And it is still a little bit of a joke. However, digital map, what you're about to see has given three speaking engagements while I simultaneously talked somewhere else last year. So it's, it's funny, but it is in a weird way. You, a little bit of a preview into what the future could be. So I'm going to just share my screen and I'll let him talk to you for the next 90 seconds. So let me share. Thanks, Steve. Thanks, Leon. Great to be here. I'm digital Matt, which to be clear means I'm the real brains behind biological Matt. He gets the frequent flyer miles. I do the actual work. I've been asked to explain our protein language model or PLM in 30 seconds first to a highly technical audience. And then again, in 30 seconds to a group of second graders. So first, the technical version, our next generation PLM is a multimodal foundational model for protein design. It fuses sequence, three-dimensional structure and interaction data, including contact points relevant to epitope clustering into a single transformer architecture. The point is not just to predict structure and isolation. The point is to rank and design molecules against a multi parameter objective function, including affinity, stability, developability, and solution behavior. And to do it fast enough to operate at scale across many programs. In practice, that lets us down select better candidates earlier, reduce experimental burden, and improve hit ranking and affinity maturation. Now the third grade version. Imagine some medicines called proteins are made of tiny letters. And those letters spell instructions for how a medicine is built. RPLM is like a very smart reading machine that is looked at lots and lots of protein examples and learned the difference between ones that work well and ones that don't. So instead of making thousands and thousands of proteins in the lab and hoping a few are good, our model helps us guess which ones are most likely to be the best medicines before we even make them. Basically, it helps us make better medicines faster. Honestly, that was pretty easy. Stephen Leon, thank you both very much for having me on practical AI and healthcare. He's also pretty sarcastic. And yeah, he doesn't age. But yeah, this thing is a new funny, what I asked Digital Matt to do before coming on with you guys today was I asked him to, you know, please come on the show and give an example of how you can tune your speech to your audience. And so, so Digital Matt talked about giving a talk to a group of technical experts on a topic we talked about last time. And he also then converted that into a 30 second spiel to third graders. And that dialogue was generated by an AI agent that I trained in my speaking style. And I did not feed at the information. He has access to the information that that we're working on and different programs and different projects. And so that dialogue was able to be generated and then fed into the system that builds both the video and the audio, which is trained on my look and voice. So that's the way that it's that is actually created. But yeah, if I had to sit and type out a script, it would take me just as long as recording it. But I don't. He can write the script. Yeah, that's awesome. Matt. So we need to start wrapping up. We've only got about three minutes in the schedule. So we'll see if we can bring in Digital Matt into this through the magic of post editing. But Steve, do you want to cover the things you wanted to cover in your section? Well, I have to be the close. You know, I know where if you can do this in one minute, that'd be great. You know, the R&D space has been complicated and has been running in a very consistent way for quite literally decades. There really haven't been any seed changes that have been this dramatic, at least not in my career's time. But in the last two years, maybe two and a half years, things have dramatically changed. How's it going to look five years from now? Yeah. So I can tell you how I hope it will look because I think it needs to. And time will tell if it actually does. But how I hope it will look is that we could dramatically shift the time and cost curve, which has been, as we said before, time stubbornly consistent and cost ever increasing to accelerate the concept of a drug to getting it to patients. And I think the way that that will look different is that a much more automated, agentic way of building all of the systems we've talked about over the last two episodes. But pulling them together to operate autonomously, human in the loop for decisions, but all of the stuff behind those decisions is pulled together in the right way at the right time, where agents are looking up, what are the right models to answer the question, not using models that are not relevant to answer the question. And then feeding that forward to the next step, the next step, the next step in the process. So I love to see all of this done in a much more automated way. And I think that's where the real acceleration will be. Yes, we're getting periods of acceleration within a function. We're getting more efficiency within a function, but to really bend that cost curve and that time curve across the entirety of R&D process, it needs to be done non-made away. And I think that that requires that orchestration of all of those agents working together. That's what I hope it will look like in the next five years. Fantastic. Hey, Matt, I just want to thank you for an excellent episode. I'm going to hand it back to Leon so we can wrap. Yeah, just you want to thank you, Matt. This was Matt Trupo, Global Head of Research and Platforms and Computational R&D at Sennifey. Two episodes, one incredible story of AI across the Triad Drug Development Pipeline. Thank you for being so open and generous with your time. I think this has been a tremendous service to everyone in the industry. And it's also been a pleasure, you know, just enjoyed the conversation very much. And I want to thank our audience for tuning in. If something in today's conversation resonated, send us a comment on LinkedIn. And we look forward to seeing all of you next time on Practical AI in Healthcare. Thank you for joining us this week on Practical AI in Healthcare. If you're ready to go beyond buzzwords and hype and explore how AI is truly transforming healthcare, stay tuned for more conversations that get us to what works. Until next time, stay practical.