newsfilter
← Latest News
Earnings Call Transcripts·

AI Panel - NOVABRIDGE BIOSCIENCES

So very good. So if we could start the broadcast. So thank you all very much. I hope everybody's had a great day talking about our Novartis and our portfolios. Thank you also for staying for this last session. We have the opportunity to discuss AI and R&D. So with me, we have Fiona Marshall, as you know, our Global Head President for Research. We Have Max Lawson, who leads artificial intelligence efforts in Novartist development, and Christian Diehl, who leads our technology and artificial intelligence efforts in Novartis research. What you see behind me is an outline of how we deploy AI at every single step of the R&D process. For those of you who have not had the opportunity outside, we have actually two displays that Max and Christian have been answering questions. So you certainly after the session can take advantage to go deeper. But as you can see at each step from discovery to lead optimization, pre-clinical and early clinical development. We work on embedding AI in every single part of the research, our research efforts, and then in development, more for this operational latency that I talked about, speeding up how we work and optimizing how we working every step of development thereafter, all with a goal to shifting probability of success and overall timelines as we outlined this morning. I thought maybe to kick off, I'll ask a few questions, But then hopefully all of you will have your questions ready, and we'd love to have a dialogue about our efforts in AI and R&D. So, Christian, maybe starting with you, when you think about the AI companies that are forming now in drug discovery, what is the alpha for Novartis in this space? I think it's threefold. So it's a level playing field, having access to compute or having access zu models. There's millions of open source models away, so that's not a differentiated factor. The alpha for a company like ours is, A, the access to internal data that you have generated over decades. So to give you a flavor, we have more than 150 petabytes of research data that we've generated in our laboratories and our trials over the years. Now, it's very difficult to imagine this, so the proxy is it's about 40 million years of 4K Netflix streaming, if you would watch it. So it's a lot of data that мы can exploit and put to the table. And having a model that is trained on open source data that everybody can access becomes really valuable if you can bring it inside and train it on your context, on the diseases that you're interested in, and train It on the knowledge you've accumulated about certain mechanisms of actions or diseases. And we've started 10 years ago to organize as if we had known the AI wave would come. Ten years ago, we started Data42 to start organizing our data and extract it from the organizational silos and harmonize it, map it to an ontology, so that we can actually traverse this data now of there's a disease, a patient, a gene, a treatment, a protein. It's all being connected, and you can put AI to work. Second, I would say, is we have accumulated also decades of biomedical research knowledge, and we have excellent researchers that, of course, know their field. They know the peers in the field, and they stay on top of the science. Because AI is very good at generating things. But you need the human judgment to understand, is this correlation also pointing to a hidden causality that we've overlooked, or is it just a correlation and it's just noise? So this human judgment that you need to apply then on the AI findings is critical. You need to be able to ask the right questions and judge the answers. Getting a great answer to the wrong question won't move the needle in drug discovery. And I think the third one is you have to have this mindset of the willingness to reimagine the work processes that you have. So this is depicted now, this slide, as a bit of a sequential process. You find a target, you make a molecule, you move it into the clinic. But in reality, where we are at Novartis is these are learning loops. So we do forward and backward, constant forward and backward translational learning that is traversing organizational boundaries. So we pick up clinical signals and we inform the way we design compounds to avoid unwanted drug effects. So think of this as many smaller and bigger loops that constantly feed information, all with the point of reducing this biological latency that Vaas talked about this morning. We're accelerating our knurling on biology. Great. Thanks, Christian. And Max, you're leading our efforts to accelerate our clinical trial work. Where are you seeing the biggest impacts today already on speeding up clinical trials? Yeah, thanks, Vaas. So as you said, a primary focus of our work in development has been speeding up clinical trails. And so we are already seeing some proof points on that. Building on the success we had with V1P, where we were able to accelerate enrollment by six months through AI-enabled site selection. We have been systemically embedding AI in the way we design our protocols, in the ways we execute the study, and in the the way that we report them. So as you mentioned this morning, on the reporting side, our content for our clinical study reports is now 90% AI-assisted, which has reduced authoring and review time by 25%. What I think we're increasingly excited about is this opportunity to do much more holistic process redesign end-to-end. And so a great example of that is with our intelligent decision system, where we've taken our trial planning process, which historically took us a few weeks to analyze a few different scenarios. It was fairly manual, even with AI assistance. And we've moved that now entirely in Silico, so we can run thousands of simulations about a potential study in minutes. And what we found interesting is those simulations are also 50% more accurate in predicting enrollment timelines, even than our prior process. So it's not just making it faster, but it's also making it more accurate, which is, I think, what we'll see going forward as well in our AI work, that we can make development even more predictive and continuously optimized as we redesign the process. Thanks, Max. So, Fiona, a few years ago, you pioneered our efforts with isomorphic labs, and now we work with a number of the players in the ecosystems. What are the types of collaborations you're seeing currently in Novartis research? So, I mean, we want to move across all of these domains as quickly as possible, and we can't just do that in-house. So we're continuously looking at who are the pioneers in the different sections of applying AI to the drug discovery and development process. And Isomorph, it was a good example of really being that first company that was able to do the folding of proteins and leverage those protein structures for drug design. So that's why we wanted to work with them. And we apply that thinking really across this whole process. So we have many different types of collaboration. Some at the earlier end are often about getting access to data, not just data that's already in existence, but also generating new types of data and particularly with a focus on human clinical data. So Relation is a good example of that because they're based here in London. They have access to London hospital tissue in diseases that we are really interested in. And they're able to get fresh tissue and use multi-omic profiling that then can feed foundation models. So sometimes you actually want to generate that type of data that's not already out there. Then there are other examples where people have developed extremely good software for predictions. Chai would be an example of that and apply those in different ways. So we want to identify new targets by accessing novel biology, optimize molecule design, whether that's small molecule design or biologics. and then obviously then safety prediction is I think one of the most I think enabling things that can really change how we transition compounds from pre-clinical into development so here we have a huge database ourself but in this case it wasn't fit for AI algorithms so we partnered with Decifex to digitize all of our slides and enable that so you can see there's really different types of collaborations but fit for purpose on to map onto all of the stages that we want to work in. Maybe one last question before I open it up to you. So Christian, one of the questions that often comes up is compute. Do you need compute in-house and models? Do you build models in- house or only take them externally? Maybe you could provide some perspective on how we think about compute and models. Yes. So compute is not a constraint for us in itself. So we have our own high-performance compute environments and we closely watch the demand of what the workloads are on these compute environments. And if it's not enough, then we scale them by adding additional GPU stacks. We can also, through our partnerships with companies such as AWS, for example, burst additional compute demand spikes into the cloud and just rent the compute capacity that we need to. On the models, we do all of it. So we partner with companies that have proprietary models develop themselves, so such as isomorphic labs, as Fiona just mentioned. We are eyeing closely the development of open source models, so on public platforms. So every open source model that is published in our area gets ingested into the Novartis tech stack and is available within 24 hours if it passes the security tests to our scientists. And we also have more than 600 data scientists and AI engineers working in research that also, of course, have the capability to create their own foundational models or deep learning models. So we are choosing the right tool for the right problem and in the best efficient manner. It may be one building on that last point, Christian. On cybersecurity, Max, as we use these tools and increasingly deploy AI, of course a lot of concerns from a cybersecurity standpoint. Maybe you could just give an overview of how we think about cybersecurity in our clinical data? Yeah, I would say, I mean, in general, what we've done is we've built that into our overall control framework that we put all of our AI solutions through. So cybersecurity, data privacy, all of the various risk factors. We take a risk-based approach in general to how we apply those controls and make sure everything is in place. With cybersecurity specifically, there are a lot of, we've obviously made a lotof investments in making sure that we've got a robust setup for that. There's also a lot fo continuous monitoring. There's disaster backup plans, et cetera. So I'd say we're well-equipped to manage that as part of our overall control framework for AI. Thanks, Max. So I'll have to open it up to questions, see what's on. Wow, a lot more questions. All right, Steve, we'll start. I like to start with Steve always. Thank you so much, and thanks for doing the session. So two questions. One is structural, and one is bigger picture. So you mentioned the Data42. Is data 42 accessible to all Novartis employees? So is the cardiovascular team able to see the data the neurology team has and so forth? And when data is found within data 42, how often is it usable versus just throwing up your hands and saying we've got to start over? And secondly, you mentioned that NovartIS has decades of data, but so does Merck, so does Pfizer, so Does Roche. Some of those data might be useful to you. So I'm trying to understand, what is this reluctance to sharing data across walls if, ultimately, it comes down to the human that you employ? So why is this data so important when it's the human who ultimately matters? Thank you. Christian? So is it accessible to all? That's the point of Data42 to enable cross-disease learning. So, for example, we have created products like our B-cell pool. where we've pooled all our insights on B-cell depletion from immunology, from oncology, and from neuroscience, and that is accessible, of course, to all researchers to accelerate the learning on what happens in B-cel depletation and how do we go about it. So without such a platform, this would be impossible. Given we're often dealing with clinical data, you have, in order to be ethical, all of the data is anonymized anyhow, but secondary analysis has to follow a purpose. You can't just dig through human data just because you feel like it. So you have to state the purpose of your scientific use case, and then you get approval to access the data that you need to. Is it useful? Yes, that's what we've been doing the last 10 years, to harmonize the data sets so that they actually compare to each other and are mapped to an ontology so that we don't start then with three months of data organization before we actually enter the data science use case that we actually want to do. Maybe I'll give some examples of where we've actually used it in practice. So one of the areas where we're linking preclinical toxicology and safety with clinical safety signals, and if we get a new drug that's coming through and it has a particular safety event, that could be either preclinkly or even a clinical event, we can actually easily go into data 42 and see, has that happened before? what was the outcome was it predicted pre-clinically and we have made some important decisions based on that either to stop a program because we see something pre-clinically that did translate into the clinic or actually go ahead because we've seen this before and actually it wasn't a problem so that's a very important part of the decision making maybe i'll just cover the sharing of data because i think we do we having been at other pharma companies there's no i genuinely no other company has the data laid out in this way uh in the with the ability to search it in the way that we do others are trying to play catch up now and it's not a trivial thing to do we do share data with other companies and i'd say particularly in the pre-clinical safety data for example um so we made the decision when we work with decipher x that that actually they would take all the data and then it would be um sort of shared so that because we we need to really work with the regulatory authorities to try and move to NAMs where we can use predictive models and that everybody is using the same models and criteria and that we get those approved to go forward. But where we think it's proprietary, maybe we're the only people that's run a successful trial in Sjogren's, we're not going to tell everybody else how we're doing that. I think that's the balance Roy's trying to strike. Naresh? Thanks for taking my question. It'd be interesting to learn where AI is creating like really truly novel outcomes so if you think about the molecules where you that have been generated using ai how many of those would not have been developed with the knowledge already inside the company and are those molecules actually any better is any evidence that the molecules are better or are they just different so we are seeing starting to see the evidence maybe fiona you want to cover this yeah so i mean real real examples first of all if you take the isomorphic collaboration but this is also applying to our in-house now methods we have targets that we have tried before and failed using all the conventional methods of high throughput screening and normal even structure-based design, not found anything. And we're now with new approaches able to find different binding sites that actually can turn out to be functionally relevant for protein function. So we're definitely at that point. In terms of finding molecules that you don't find before, what AI is particularly good at when you do the generative chemistry is it actually does take you into a wider area of chemical space that is not normally explored within a chemistry program. So would it eventually have found, would you eventually found it? Probably, but maybe you wouldn't have got there for, you know, five years, whereas you're actually finding and exploring a much wider areaof chemical space within a few months in this case. So I think it allows you to explore wider space, find better molecules quick, more quickly. It made me feel also the example on a novel target, let's say in neuroscience, what we've seen? Yeah, so in the target identification where we're using foundation models that bring together data sets from both clinical, preclinical, cell-based, multi-omic profiling, in the neurodegeneration space, then we've already got three novel targets that are actually advancing into our later stage programs that have gone through this process. What's very important, we believe, is having really robust lab-in-the-loop profiling assays so that you take the output from the AI models into relevant either cell-based models or then later on even in vivo models where you can validate what comes out of the AI modules because actually still a lot of the things that come out just don't translate into even the cell models, let alone into the clinic. Maybe just a reminder, state your name and your firm as well. Simon? Thank you. Simon Rick from Rothschild. Two related questions, if i may most of the effort that ai players outside pharma are putting into this into drug development is at the very early stage target identification because it's the most accessible you don't need labs you don'T need to be on clinical trials but there was a good paper in august suggesting that's not really where it should be applied that actually looking across the value chain the biggest impact is in reducing phase two clinical failure rates so your thoughts on that would be interesting and going back to that first question in terms of management of expectations if we look at alpha fold which is sort of held up as the poster child here what they did was truly amazing but what they really did was machine learning on an extremely high quality data set of crystal structures has that created false expectations leading to comments from people which aren't named that have said that cancer and alzheimer's will be relics within a decade. Thanks so much. Yeah, so maybe just broadly, I think it's worth noting as well, while these are our pioneering projects here, we also basically give AI tools across the company. And we are seeing already compounding effects from, as I mentioned this morning, our biostatisticians, our clinical reviewers, all having incremental gains from using these models to speed up their work. So that's not captured here, and that's something that's broad-based, actually getting deployed quite rapidly across Novartis. But maybe to break up the question first on the phase two probability of success, Krishnar, Max? So I know the paper you referred to. It's a good paper. It's an excellent paper. It's good paper, and so we're trying to, we're working on increasing the probability of successful. So one example is this predictive safety use case where we're try to predict early on potential toxicities. So it's basically turning the problem around and say, well, I can't predict success, but I can potentially predict potential failure. And if I can predict this earlier without going into the clinic and finding out then, then I can avoid this failure. This is something we call safety by design. I can integrate this into my compound design process to come up with safer So that's one way of driving probability of success up. Finding new targets is, of course, exciting. We have many diseases where we still lack, collectively, the right targets. Will they pan out to be successful? We'll have to go through the clinical trial to assume this. What's also driving probability-of-success, and I'm sure you will see the impact that will make our ambition real, is that we're now able much more to simulate different versions of trials in different subgroups of patients so that we actually recruit the right type of patients that will show the response that we're hoping for rather than having a bunch of non-responders, which will also drive phase two successes up. Maybe one other element there. I think our endpoint are going to get much more sophisticated with the power of AI, and we're already seeing these effects in neurodegeneration and neuromuscular diseases. The sophistication of the endpoints should help us better discern real impacts on patients. Now, maybe on the alpha fold part of that question, Fiona. Yeah, so, I mean, you can see by looking at this slide that to actually get to a medicine, you have to go through this entire process. And I think everybody in the room understands that. So, you know, just fixing one part, in this case, generative chemistry to make much better molecules is an important part. Of course, if we have better molecules, then they should fare better going through this process. But actually, we were already pretty good at making molecules already, and we've made some remarkable molecules. But I think that there is generally an underestimate of what it actually takes to convert those molecules into medicines and the complexity of really understanding human biology, which we still are not very good at, even in the context of AI. And in terms of understanding human biology, which is what you've got to do to cure all diseases, we actually have a lot of data on the diseases we've already been good at treating. And so maybe you can use AI to find some more drugs for those diseases, but I think when you get into diseases where you've never done a trial before and you know very little about it, that's going to be extraordinarily difficult for AI to crack. The paper Simon refers to is In Nature, I think it was the end of August. It's a review of drug discovery approaches, a very, very, I thing, strong paper. Please go ahead. Wait for the mic. Rajesh Kumar from HSBC. Two, if I may. Appreciate the value of data. It obviously is your superpower. Does it apply to generative search? Does it imply to machine learning? I would imagine data is more powerful when used in machine learning techniques, which have been there for a very long time, right? So what I'm trying to get down to is in terms of AI is a very large and compassing word. And a lot of these elements have existed even at a time when, you know, I was a computational chemist 25 years ago, right? So it's a very old technology, so to say, but there are certain parts which have accelerated. So, can you help us understand this whole data thing you've done on top of your AI? What is new? What is it that is recent and that is going to change the things? Second part of the question is related to, you know, the point about that Nature paper, a crystallographic point, is we all talk about improving the probability of success or whether it stays the same. As I manage my equity research team, I often find that the ideas AI generates lack originality. They're often very comparable to what, you know, wishy-washy undergraduate thinking would produce. Depends on the undergraduate school. I'm talking about myself, right? So I'm not that bright undergraduate. So I often tell my analysts to start with that data and think on top of it rather than use that as an end point. The problem we've got is we have discouraged a whole generation of scientists to become crystallographers today, right, because AI will solve the protein structure. What happens if your organization keeps on seeking productivity and actually it replicates what a lot of other organizations are doing? You all get crowded. Probability of success goes down. Who will be responsible? I will give that to Fiona, that second part of the responsibility. But I think first on why the data matters. Christian? Yes. So, if you look at this slide, these are not LLMs that we apply in many of these fields, right? And so, you're right, we have done computer-aided drug design and chemistry for many years, and there are machine learnings and deep neural network approaches that we take. But there are other things that are brand new. So, in TargetID, for example, we do in silico-pertubation models based on the underlying technology of the LM, the transformer, where we apply transformers to genome sequences and cells. and that leads to new insights, potentially new insights. It leads to correlations that then our scientists have to judge. That technology at scale wasn't available two years ago, so that we didn't do for the last 15 years. So that has changed. Foundational models are getting better. There are newly ones invented. We watch the market closely. We read the same papers that also the tech companies read that come out of academia. So I think we're state of the art. There is rapid improvement of the capability of these models. You see this, for example, also in the de novo design and biologics that a field that was nascent a couple of years ago, now it's starting to pick up and the model predictions are starting getting better and better. So the better the prediction of these models, then the more we can start to actually use them and put them to work and know which molecules actually to make or which targets to pursue. So that has changed. It's a continuation. So we didn't start with AI two years ago or three years ago when we heard about ChatGPT. That's for sure. We have done machine learning and advanced analytical approaches for decades. And the data lake that Christian refers to, Data42, we started building that with Palantir in 2017. So that has been there. I think the speed at which now we can link some of the AI-generated insights that Christian just mentioned and then go into the data Lake to in silico test them and then bring them back. So it's kind of compounding, but to your point, some of these things have been in place. I think AI is just now accelerating our ability to leverage all of these tools. I mean, similarly in Max's space, Max, you want to comment? You and I have been trying to speed up clinical trials for a long time, but agents have really helped, right? Yeah, and I mean I would also just add to the data comment that I think, you know, part of the difference of getting a good answer out of AI versus a bad answer is the context you provided. And so I think curating all of this data and properly linking it and providing high-quality data into it is part of the differentiator. I think in the operational space, a lot of this is operational data that we would add into this, our own data, also data that we would license or get in from the outside. And so I think providing all of that kind of context into some of these operational use cases is part of what drives a lot of the quality of the outputs that give people confidence on the answers that then allow us to take decisions quicker and move things faster. I think just to kind of go back to the Data42 example in the context of some of our evidence generation work, you know, our teams building on top of Data42 were able to make the feasibility analysis part of the evidence generation process 20 times faster by adding AI onto the real-world evidence data that we have in Data42. And you add that to the protocol generation part of it, which also relies on our historic protocols in that space, which were 40 times faster. And you see this sort of end-to-end compounding effect of building on our data to create an end-TO-end workflow that's radically faster than what we were able to do before. Maybe Fiona on the talent question, and how do we keep people to be creative minds in the age of AI? Yeah, well, actually, I haven't seen so much of a problem. I think that this is actually stimulating even more creative juices to be flowing, and people are really excited about the possibility. It also frees them up enormously from things like making PowerPoint slides, which we had incredibly smart people doing. so it frees up time to actually you know do have this these more creativity ideas and I think also it enables more people to do different types of things that they weren't able to do before so for example medicinal chemists can now do computational chemistry that they wasn't able to do and these so it actually trains their mind in a different way so I'm quite excited about how we're seeing talent develop now in ways that they didn't do before. I mean Fiona mentions it in a company that generates a lot of slides, the speed at which now we, that our people can generate perfectly formatted slide decks is a tremendous productivity unlock. And so I think the heart, you'll start to eventually see it, I'm sure, in the numbers. Back, Emmanuel. Emmanuel Papadakis, Deutsche Bank. Thanks for taking the question. Just to understand on this slide, it sounds like there's no intent to build a unique proprietary go-to holistic internal model you're going to rely on a multitude of internal and external models going forward depending on the problem and the particular point in which any researcher or internal individual is trying to solve a particular problem is that correct? And then the second question on that is the extent that you're gonna rely on an ongoing basis on external models how are you controlling control and security of the data flow? Do you have folks inputting important internal data or interrogating external models with questions, et cetera. I think you'll know what I mean. Yeah, we've put a lot of thought into being very thoughtful about what data we allow to be used in other models and how they get trained. Christian, do you want to just give an overview? Yeah, so we're generally very hesitant to let externals train on our data and then run off with a model with improved weight. So what we usually try to do is we'll try to find an agreement that we bring the model behind the Novartis firewall and then train it, and any fine-tuned model on Novartis data has to then stay with NovartIS and cannot be sold to our competitors. On the question of world model, that would be nice, but that's why we chose the name 42. So for those who read the Hitchhiker's Guide to the Galaxy, there is such a world model. And after consuming endless resources for generations, it spits out the answer to life, universe, and everything is 42, which is probably correct, but it's not very useful. So I would say, no, there's no intention to build the all-encompassing world model. We are building biology foundational models that are disease-specific, right? So that, to understand and to unlock the mysteries, what are the ultimate disease drivers? And I think that is the better way to be more targeted. I think Seamus in the back. I'll just try to move back to the front of the room. See you guys. Thanks, James Fernandez, Guggenheim. So I guess in the context of recent data, how can you basically utilize AI to interrogate, you know, VHOT or to interrogation endpoints that may not be super reliable? I just think it's interesting in the context of, you know, being able to dial back on certain endpoints. So I'm just interested to try to understand, do you see opportunities in the wake of data like we had from the avidity side of things where there's a lot of learning that's gained on endpoints and you can actually bring that to regulators to educate them better? Thanks. Yeah, I think broadly speaking, I'll speak in general terms, we are using, you machine learning to really rethink endpoints that involve human judgment in terms of, so if you think about, take the example of VHOT, it's a video, it comes to central adjudication, a trained reviewer decides based on their judgment, when did the hands start to open and when did they fully open? And if you can use machines to make that an incredibly consistent approach that with incredible precision and optimize, that obviously gives you improvements. And that can be done across many of the endpoints we historically have used in studies like that. And there are many of these endpoints. You know, we had tried this when I was running development over a decade ago for some of the multiple sclerosis secondary endpoints, and the technology just wasn't there. Now it's there, and we can do this with incredible speed and precision. So I think this is going to lead to us improving endpoint step by step. We will have to get the regulators to get comfortable, but as long as we can show these are validated, I think that should be doable because I'm sure they would also prefer not to see the huge variability you see from humans doing this. The other thing is we're actually enhancing data quality, and the particular area I think is in the translational AI where you have measures such as imaging, MRI data, cardiac signaling data so you know atrial fibrillation is one of the areas we're putting a lot of effort into so we have gone back and used AI to analyze all of our EKG data both in healthy people as well as in people that may have atrial fibrelation and there's actually a lot of those in there and so this is really helping us to look for the signals that we would be detecting or the size of the trials and actually plan on using AI enhanced uh ekg to take forward some of our atrial fibrillation portfolio james i think it's coming thank you uh james gordon from barclays he said that the slide looks really comprehensive but i know there's some disclosure differences between companies and i think you've been more open than some other companies about what you're up to so in which ways do you think novartis approach is most differentiated versus like what your large-cap pharma my competitors are doing? Is it the data organization? And where would you rank Novartis now? And then how durable do you think that's going to be? So you've helpfully put this up there, but do you thing everyone else is going to quickly copy it? Or are there things that you can do that are going to give you a durable benefit? Yeah, I can start and my colleagues can jump in. I mean, I think a lot of the things up here are not based on technology that's particularly proprietary, but it's going to be your ability to have data that's well organized and orchestrated that can actually be leveraged by the systems, which is not at all straightforward. And I would say we've, because basically all three of us, we tried to do this a decade ago, and we just didn't have the right technology. I mean, if you could rewind one of these back to 2015, you can find me talking about many of these things, but we just weren't ready yet from, but We're working on the data. And so I think we are very well organized relative to our peer sets on operational data and scientific data. That doesn't give us a forever lead, but it does give us a lead. And then our ability then to scale because we have that data, so we can use that scaling to have an impact across our whole portfolio, it compounds, and then you have a lead for certain. So I think that's how we think about where we're differentiated. And then, of course, there's, as our team already articulated, there's going to be areas where we have very deep data sets. I would argue that the areas of cardiovascular outcomes trials over our history, the number of cardiovascular outcome studies we've run is on the high end of the sector. And having all of that data in-house means we can mine that data in ways that others can't. There are going to be other disease areas where we have very deep data sets, which are really going to Be hard, because it will cost you billions of dollars in a decade to recreate that data set, just because you have to run the studies. So, I mean, those would be kind of my high-level thoughts. Christian or Max, anything you want to add? No, I would also say the moat is around the diseases we have been in for many years and we have organized the data. And I would for once disagree with you that we can, that somebody else can replicate data sets because it will be unethical to, even for us, to rerun our own clinical trials. No one would allow us to do this. So this is really proprietary and unique, right? So, and I think it depends on the capability of each company to say in the areas we're interested in. We're the only ones that, for example, have the Kantos trial. There's no one else and no one can replicate this data set to query it. Other companies may have other data sets that are unique to them. And then the willingness and capability to embed AI across these processes and then redefine how you run your research and development processes and discovery processes. Every company has the ability to do this. I think we're way ahead in thinking about How do we rewire this from a sequential process, as I said, to these lab-in-the-loops and learning-loop concepts and so on? Yeah, and maybe the last point I'd raise is how Fiona is actually organizing the, you know, in areas like chemistry around AI, which will then, I think, really embed it in how we work, right? Yeah, I mean, I don't know. I think there's a couple of things for me. First of all, other people can work with these companies. They are. But first of all it's how quickly you can also adopt the technology and actually convert your own teams to running in this way. And that can involve structural changes, hiring new people, uptake of the technology into the work that we're doing. That's the first thing. But I think the other thing also just comes back to still I think The Unique Advantage is the deep knowledge we already have and the questions that you ask. So everybody may have the right tools, but if they don't know what the problem is to fix, they're not going to ask the right questions. And so even when we've gone to these different companies like Generate Bio or Isomorphic, we're coming with problems to say, can you get AI to solve this? These are problems they didn't even think of or know was even a problem or things that they wouldn't. So I think that's what gives us that real advantage, always has done. And AI is just helping you to answer the problems we can conceive of more quickly. Flora? Thank you very much. Good afternoon. Florence from OdoBHF. Two quick questions, please. First, where do we, we should see the low-hanging fruits using AI? Could it be in manufacturing? Could it be commercial or research or development? So it would be interesting to know where you could take easy to implement measures to improve productivity. Second, when should we start to see an improvement of the probability of success of the development of the new drugs? Of course, without waiting for the entire cycle, could it be in two years, three years from now? Do you already have products entering clinical trials now that should read out in one or two years that you could say, well, we have properly or better designed our trial and then we have a higher probability of successful? Yeah. So I'll take the broad – you asked really about near-term productivity gains, right? And then we can turn to the colleagues on POS and speed. So, I mean, in R&D, our goal is output, R&B productivity. We're not looking at R&P as a cost area to optimize very much. Can we find more medicines more quickly? I think in the near term where we're going to see a material impact is in manufacturing where – we didn't talk about it this morning with Stefan, But I think we're one of the leaders in the speed at which we're introducing automation and robotics into our manufacturing sites, doing it at scale. We don't talk a lot about the COGS line because, generally speaking, on excluding royalties, we believe we're best in class in COGS. And I think that's because of the pace and speed at Which Stefan and his team deploy these technologies, fully automated in areas like RLT. It was on the slide earlier. So that's where we're going to see an impact. Second is continuing the journey on our third-party costs and using AI to reduce our reliance on third parties, about $11 billion in third-parties spend. Over time, you know, there could be a significant proportion of that that we can address using AI tools, and we're starting to scale that. And this year, we already exceeded our internal expectations for the impact of AI on that. And then I think there will be, we hope over time, increased productivity for the top line as well. So if our hope, and I think all of our peers are doing this, the question, again, is who will scale it best and deploy it best. Can you get your sales reps to be and your commercial frontline teams to be more productive because they have better information, know where to go? We've been chasing this as an industry for a long time. But now when I look at the tools we're providing, it's just a huge step function from what we were talking about five years ago. And most importantly, the salespeople actually want to use the tools that we're giving them. That was the significant challenge we had in the past, that we would roll these things out and nobody wanted to use them because they weren't that helpful. Now, in terms of when we'll see first POS changes, Fiona, your view? Yeah, so, I mean, you know, Vaz, you already showed at the beginning some of the productivity changes that we've had within biomedical research, and I think one of the key ones that has been AI-enabled is that reduction of time it takes to start a program to get to a candidate, which was an average of three to four years. We've now got it down to less than two and a half years. A lot of that gain has come from using AI-generative chemistry approaches. And we actually now have, you know, we're finding what we call AI-first, where it's really AI-driven programs get down to one year cycle time. So these are already measures, and we're actually carefully collecting these because we want to bring this together to actually show you the hard comparative data. But we can see it already. So that's one we're definitely having a real impact on. and we also now have around eight or nine molecules going through dose rain finding toxicology that have come through this generative chemistry predictive safety process so that will soon give us enough to actually calculate whether this is having an improvement on the safety attrition rate i would say we would have that within a few years that sort of data but then you know on the clinical trial improvements i mean comment i'd say we're we're already seeing these sort of weeks and months savings in different sections of the process, which we look at in composite cycle times. We're deploying AI broadly across the entire portfolio, so we'll continue to make these incremental gains, and we'll see that broadly across the whole portfolio over time. Yeah. I mean, I think, and we'll just keep collecting the data. Look, everybody's going to make various predictions. I think the truth is nobody knows, right? Because with these exponential technologies, it's hard to predict. The key is that as long as we keep scaling at every step, suddenly the compounding should happen and we'll see the bigger gains Kerry Thank you very much Kerry Alford from Baramark, just going back to something you just said about a reluctance to use some of these AI tools internally historically, how do you encourage the adoption of AI internally, does it just improve in time as the functionality improves Do you use certain targets, incentivization across R&D, but elsewhere as well? So I think the observation has been that you want it to be truly demand-driven because the technology is so good that people just want to use it. And I think that's what we're seeing in most of these cases is when we really build something that helps people do their work faster. As an example, if you think about our clinical safety reviewers, I mean, the speed at which they're adopting new tools is extraordinary because this is saving them months of work or statistical programmers. And so when we have a good tool, people will use it. Usually in our company, when we set these top-down demands that people use it and that kind of management is forcing you to use stuff, it's the same way all of you feel if somebody forces you to do something. You have to use some tool, right? So I think we want it to be that we have great solutions we're rolling out, and then the demand comes. Rather, it's an outcome target that I'm thinking about. I mean, how can we really show we're improving POS, we're improving timelines, we'RE driving productivity, and we'RE improving sales? I want to be more focused on outcomes than things like token utilization or number of people using an application. And I think Fiona consistent as well, right? Yeah, totally. That's a culture across Novartis. But, you know, I think the other thing is we have rolled out incredible training programs for everybody to really upskill in AI we call AI fluency and so we have the broadest range of tools I think I could imagine and people really know how to use the tools so that that's not a hurdle to entry yeah we've rolled out AI fluency now to the whole company and there's three tiers and step by step you know basically trying to get everyone up up to speed Richard in the back thanks uh Richard Voss at J.P. Morgan just wondering about the implications of this for for scientific collaboration uh you know data is presumably going to be quite critical so how are you thinking about sharing data maybe phase one phase two or even preclinical data at scientific meetings and how that might actually affect um you know longer term collaboration and for the industry if you know you've been thinking a lot about this publication challenge yeah yes i mean actually it's we're we're collaborating extensively certainly i mean the very first one on there is an academic institute, the Helmholtz in Munich. They're generating a human cell atlas. So this is more on the research side of it, but they're doing all the AI and the dry lab work, and then they're collaborating with us because we can do the wet lab work. So it's this, it's complementary expertise is coming together in a collaborative way. And the same is actually happening, I would say, in the clinical environment. I mean, One of my observations, and I'll hand over to you, is a lot of the hospitals, big hospitals that have a large amount of data, actually realize the value of that data. And they're very open to collaborating with pharma companies to share the data and to allow us to use that data to really discover new medicines. So I think this has actually increased collaboration in my experience, but maybe from your point of view, from the clinical trial data, and then maybe a bit about publications. because we can get to meaningful data, I think, more quickly, which also stimulates publications. Yeah, I mean, I Think there's a lot of opportunity to collaborate on being able to access a lot of that data also for topics like finding patients in clinical trials and things like that. So as you pointed out, Ithink there's a lot more openness to be able to leverage that in our clinical trials. And so we're in discussions on a number of those areas to basically be able to accelerate enrollment through better patient targeting. And I think it's a great example. In terms of publications, we're being thoughtful. I think we want to continue to accelerate scientific discovery broadly. And many of our researchers take pride in publishing, but at the same time in highly competitive areas. And because I think, Richard, your question really gets to the point where AI can collapse the time that a fast follower might be able to catch up. We have to be a lot more thoughtful. And so we're putting that in and putting those considerations into our publications. And there are situations where we disclose very late. I mean, GHZ today, we took all the way through a phase 2B study before we really tried to disclose anything about what we were trying to do for very deliberate reasons. Other cases, when it's already competitive, we can be, I think, more open. But certainly that's something we're actively considering on when we disclose what on our portfolio. Some other things on this. I mean one of the things is we're using AI in the generative chemistry space to really build up estates of patents around novel compound series so that we have huge layers like an onion of protection, way more than we would have had before in the past to really try and protect intellectual property. But certainly in the area of target identification, because for me, new target identification is the one that we really want to keep close to close and not talk about those until we've taken them a long way through and got some validation. So we're out of time, unfortunately. We will, Max and Christian will be available at their sections. It may be even more impactful when you have a drink, as you actually are able to now listen to them. So I hope you'll stay for the reception. I want to just close and thank the Novartis colleagues at the back of the room for engaging and working so hard to prepare for this meeting and engage with all of you. And then most importantly, thank all of the investors who spent a day with us, as well as those online. I hope you take away our conviction on the next five years, our conviction on the five years after that, and the quality of our people and the science at our company. So thank you again.

Company
NOVABRIDGE BIOSCIENCES
Keywords
NOVABRIDGE BIOSCIENCES