Note that the webinar is being recorded. I would like to inform you that all participants are in listen-only mode during the call. After the presentation, there will be a question-and-answer session. You are invited to send in questions for this throughout the entire session using the Q&A functionality of Zoom. In addition to that, you may also raise your virtual hand to address your questions verbally. For participants joining via phone, to raise your hand, use star nine on your phone's dial pad.
When you then get selected to ask your questions, please follow the instructions from the phone and press star six to unmute yourself. During the Q&A session, we'll start a short feedback questionnaire. It will appear on your screens as a pop-up window and includes some single or multiple-choice questions. We'd highly appreciate it if you could complete the survey.
One last remark: if you'd like to follow the presented slides on your end as well, please feel free to go to roche.com/investors to download the presentation. At this time, it's our pleasure to introduce you to Bruno Eschli, Head of Investor Relations. Bruno, the stage is yours.
Thanks a lot, Ingrid. And can I have the first slide, please? Thanks. So welcome to our fourth IR event on the topic of digitalization and our last IR event for 2024. Contrary to the events which we had in 2020, 2022, and 2023, where we presented use cases all along the entire value chain of pharma and diagnostics, as well as on the group level, today's event will exclusively focus on computational biology, big data, AI, machine learning, large language models, and early drug discovery.
The decision to focus this year's event on early research is based on the feedback we got from investors, but also on the assumption that the most significant progress for the life sciences industry is to be achieved exactly in these areas. Let me quickly take you through today's agenda. We will have one speaker with us today.
It's Aviv Regev, who many of you will know already, and who joined Genentech in 2020 from the MIT to lead the Genentech Research and Early Development. Aviv will provide us with an update on gRED's Lab-in-the-Loop concept, providing you with many examples on how computational approaches start to make inroads into every aspect of early drug development, from target discovery over molecule design and molecule selection all the way to the clinic.
After her presentation, which is planned for 60 minutes, we will have 30 minutes of Q&A. So overall, the event is scheduled for one and a half hours, and for the Q&A part, we will also be joined by John Marioni, our senior vice president and head of gRED Computational Sciences, who some of you might know from previous investor calls, who will also be happy to take your questions.
Just one housekeeping item, which already was mentioned by Ingrid before. We will have during the Q&A session a short 10-question online survey displayed. We would be very pleased if you could provide us again your feedback. With that, I would like to hand over to Aviv. Aviv, please.
Thank you very much, Bruno. A pleasure to be here today. And our goal is indeed to tell you how we are using the Lab- in- the-L oop to revolutionize drug R&D. And the first question, of course, is why do you need to do this?
And so, as you all know, drug R&D is an incredibly challenging process, and it has a very low success rate overall in the industry. So in aggregate, across the industry, around 90% of programs fail in preclinical or clinical stages. That is an enormously high number, and we would really like to modify it. So for us in Genentech and Roche, we want to ask, how can we substantially shift this in order to enhance R&D productivity?
For this, we want to look at every step of drug R&D and ask, what might be the new scientific ways in order to make it better? And it is essential to take a scientific perspective on this because if we continue to pursue only previously established scientific approaches with only incremental scientific changes in each of the key parts of this elaborate process, there is no reason for us to believe that anything would help us in our mission to fundamentally improve PTS cycle time, cost, or volume.
We have to have new science as part of the R&D equation. And so often, when you have hard problems like these, everyone thinks that there's one thing, one silver bullet that will make all the difference. Some people will tell you that it's all about the core expertise.
That means the experimental systems, the synthesis in chemistry, the human biology, the lab models that we use, and so on. Other people will say that it's all in the data. The data is our biggest leverage. Big data, high-resolution data collected by very sophisticated methods, this data will save us. And these days, it's also very popular to say, no, no, no, it's all in the AI.
AI will save the day. AI is your lever. Just bring in AI, and all your problems will be solved. Well, I want to claim that they're all a little bit right, but they're all a little bit wrong. They're right in the sense that each of these parts is absolutely necessary. But they're also wrong because none on its own is sufficient.
Not just the experiments, not just the human biology, not just the data, and unfortunately, also not just the AI. The key is actually in the way that we can put all of these together. And that is what we call the lab in the loop. So at a very basic level, this loop is a very simple idea. Say you start with an experiment.
You collect data. Ideally, you collect data with modern approaches that give us this data at both high resolution and at massive scale. You use the data from this and prior experiments in order to train a model. And then you use the model that you trained in order to predict the next set of experiments and iterate. And as we iterate like this, through these iterations again and again, ultimately, we reach our goal.
And what I'm going to do today, what I'm going to do in today's presentation, is show you how this actually plays out in practice across each of these components of drug R&D: target discovery and assessment, lead identification, lead optimization, and clinical, and even a glimpse into our clinical work. Now, as we do it, we're going to keep in mind this equation because we're going to hit on multiple parts of the R&D equations. We're going to affect its volume, its PTS, and its cycle time, sometimes multiples of those at once. And so I'm going to start with target discovery and target assessment.
In target assessment, we want to use different kinds of data and information from lab experiments and from patient samples in order to develop models and maps that would help us identify the genes and pathways that are causal based on human genetics in the disease onset or progression, to understand how these changes affect cell states or cell behavior in disease, to figure out from them the main drivers of the disease, to understand how those might work clinically in the clinical context, to understand what disease biology remains unaddressed as an unmet need given the treatments that already exist with current therapies.
And finally, we want to put all of this into some hierarchical view, give or take, so that we understand the importance of disease drivers and how they relate with each other. We call models that actually provide all of these answers tied together for one given disease a pathobiology map, and herein lies our first AI challenge because biology is multimodal.
So just like a movie is not just made of its script, but actually we have stills, and we have video, and we have audio, and we have a whole media campaign, so is biology. It is not just about the DNA, but we also have gene expression and cells and tissues and entire medical records all the way through. In the world of AI, we handle this through multimodal models. For example, you might be familiar with multimodal LLMs, which learn the association between two different modalities. In the case of multimodal LLMs, it can be text and images, and as a result, once we learn a model like that, we can then manipulate it.
For example, if we flip the word green for the word blue, the plants in the picture are going to be colored by the color blue. And analogously to this, I'm introducing our first model for the day. This is Nona. And it is one of our multimodal LLMs for biology. In the case of Nona, it connects two modalities. One is DNA sequences, and the other is single-cell expression profiles. During training, Nona learns the association between letters in the DNA and cell-type-specific RNA expression profiles.
Just like you might mask certain words when you're training a multimodal text-image LLM, here you would mask certain types of information, for example, on the RNA side or on the DNA side. After we train Nona to learn the correct DNA sequence that corresponds to specified RNA expression, we can use it and apply it to a variety of tasks.
So for example, we use the Nona model in order to predict the cell-type-specific expression that would be associated with a variant in regulatory DNA. Why is that important for us? Because most of the loci that are actually associated with the risk of common complex diseases like IBD or COPD or Alzheimer's disease, most of these loci are actually in regulatory regions, not in coding regions.
And it is hard for us upfront to, first of all, identify where the causal variant is, but more importantly, to actually understand what this variant would do. Unlike a coding sequence where we have the genetic code where we know the change in amino acid and whether it would be deleterious or not, this is much harder with regulatory regions. But a model like Nona actually becomes, in a sense, the regulatory code.
And so for example, this becomes very valuable in early target assessment when we start with human genetics data and we want to tie it downstream to the genes, the pathways, the cells, and the tissues that are affected by this variant. And so this is an example of just one pair of modalities along that stack. But overall, we assemble multiple data types with multiple models in order to build one pathobiology map. So here is an example of the first pathobiology map that we constructed.
This is in IBD, one of our end-to-end diseases in our pharma strategy for immunology, one of our five therapeutic areas. And so it includes not just the genetics and expression data that I illustrated to you, but also other types of data. For example, histopathology data from patients.
Here we can leverage the unique and rich data that we have from our etrolizumab trial, as well as data from preclinical models that we run in the lab. The output that the map provides associates specific genes into pathways and the pathways into cells and the cells to each other as they underlie the disease pathobiology. That's not the end of building these maps.
We can then add additional models. For example, we can have models that address the question of predicting ligandability or predicting safety, all of which are things that we need in order to relate to the feasibility of this target. Finally, this would lead us to the targets that we would want to enter into our portfolio. How does this play out in practice?
If we look at IBD and the COPD, these are our two end-to-end diseases for the immunology therapeutic area in the Roche Pharma strategy. In gRED clinical pipeline today, we have two phase II assets in IBD, and we have four assets between phase I and phase II for COPD. After building our pathobiology maps for each of them, and these are real data even shown in quite abstracted form, we now have between eight targets in COPD and 12 targets in IBD in the target assessment phase, as well as another 12 and 4 targets in lead initiation, LI, for the two indications.
So in this way, instead of being limited by specific hypotheses that are generated by a scientist that is thinking about one gene or one pathway at a time and is working sequentially, first I work on this, and then I work on that, and then I work on this other thing, and you would look at just one cell type or one pathway at a time based on your best guess.
We in one fell swoop have this comprehensive map for each disease. And then we can pursue it in an informed way, making our choices not just one at a time, but also based on how they connect together. And as more data gets collected, we can, of course, update this map continuously. Now, these two maps were built with multiple individual models, multiple computational tools, and by a team of computational biologists and disease experts.
This meant a tremendous amount of work over a period of time. But we want to be even faster than that and take more extensive actions. And for this, we can bring a second level of AI. And that is in the form of what is known as an autonomous agent, which can perform tasks and make decisions on its own without needing continuous human guidance.
So what can an agent do? An agent can understand its environment. It can make decisions based on that understanding. It can take actions to achieve specific goals. And very importantly, it has memory of past events and outcomes. So if it did something and something happens, it actually remembers it. Agents have existed in computer science for many, many years, actually for decades, but they weren't particularly successful.
What was really the turning point for agents was being able to equip this autonomous agent with a large language model inside it. And that significantly enhances the ability of the agent to understand because it can now have this LLM-style understanding of text to interpret for the same reason and also to generate natural language. And that makes it more versatile and more effective with many applications, as well as more effective with interaction with human users.
And as a result of that, the agent can achieve very complicated tasks or workflows. It is going to break the task into smaller pieces that can be accomplished, and it's going to be engaged with multiple tools and foundation models in order to perform these tasks.
At every step, the agent is going to reason with its LLM in an LLM-ish way about the outcomes of its actions, and then it's going to take another action without any explicit guidance or pre-coded instructions from a user. This doesn't mean that users are not there. An agent can also stop and interact with a user, and the user prompts the agent to actually do its first action.
And the agent can do it at its own discretion. It can stop and ask the user, but the user can also guide it proactively. It turns out that you can expand it also in different ways into entire orchestration systems where you have multiple agents that are orchestrated by one. You can think about it as the spymaster. And now we're going to see how this actually works for us in real life.
And so in order to help us build and explore pathobiology maps and perform many additional analyses in the biology domain, we developed our Braid agent, which you are going to see now running in this demo on an illustrative example in the domain of inflammatory bowel disease, IBD. Our agent is connected to a whole slew of foundation models. It just made its first plan, by the way, to a whole slew of foundation models.
They include Nona, the model I described to you now, Scimilarity, MapLib , lead models that I will describe later, and additional models that I did not describe today. It just finished doing a gene expression analysis, by the way, that it decided to conduct. The user is asking it another question. It is also connected to all our project data, so it is able to run those analyses on our data.
And it has huge access to tools and data that a typical computational biologist would have. It can go to anything on the web. It has a knowledge graph. It has all of our internal analysis tools. It might have access to our clinical trials. It could look at electronic health records. It is even connected to the competitive intelligence database, which you will see it actually asking a question just this moment.
Is there any drug in active development targeting this? It pulls it out, and it gives you information about its development status. So as you can see in the video, the user can interact with the agent, but the agent can also address very open-ended requests, pursue chains of many tasks, make decisions on its own what it's going to do next. And this makes building and querying maps for targets much easier.
It also makes a ton of other tasks not just easier, but much more open-ended. Many people would know how to do some of these things, and they might know how to do them better than an agent, but they won't know how to do that big variety of things. That variety really opens up possibilities for them. Okay. So to help us, and so typically in drug R&D, once we identify a target, for example, through these pathobiology maps or just based on the knowledge of a particular researcher, the next step that will come is that we will separately identify a lead compound to modulate it.
And so the next major approach that I want to describe to you that we're taking is actually to combine these two stages by using phenotypic screens and AI in order to identify leads without necessarily specifying the gene target upfront. For example, we might want to study a pathway or a cell type in a disease, but not just one target. And yet we would still want to identify targets and find leads at the same time. And so this really compresses two drug discovery phases into one, and I'm going to focus on it next. I'm introducing to you our first loop of the day.
There's going to be several. In this loop, we are going to start with cells. We're going to screen them under hundreds of thousands of perturbations. These include both genetic perturbations like gene knockout or knockdown or overexpression and small molecule perturbations.
And then we're going to collect very rich profiles on these cells. These include RNA profiles of individual cells and images of individual cells. From these, we can learn a model that relates the perturbations to their effects and to each other. Then we're going to query this map for predictions, which simultaneously finds targets and small molecule hits that generate the desired effect on the cell, and they become the starting point for drug discovery.
One of our major efforts in this area is our collaboration with our partner Recursion. For example, in cancer, we conduct two loops in parallel. One loop, which you see at the top, is based on cell images. Those cell images are acquired by Recursion. Another, which you see at the bottom, is in data acquired in genetics. Here we use the same system, but we're collecting single-cell RNA profiles.
The Recursion data includes both whole genome knockout and hundreds of thousands of small molecule compounds. So we can reconstruct things with AI like the compound knockout relationship. The genetic data includes a genome-wide perturbation screen in the same system. So it provides for us this in-depth reconstruction of each gene's function and of entire circuits, how genes relate to each other, and the complex molecular phenotypes that they create in cells.
By having both of these views, this multimodality and two kinds of perturbations, two genetic and small molecule perturbations, AI can generate for us a map that indicates which small molecule perturbations may be associated with which cell phenotype and for which potential targets they act, based on the information from the knockout. This year, we collected with Recursion also the first map in our main area of collaboration in neuroscience.
The first neuro map, as we call it, includes a genome-wide CRISPR knockout screen with images along with disease-associated perturbations. It required scaling screens to a trillion cells, and just for a sense for you, you are made of about 37 trillion cells, so this is about one-thirtieth of the cells of an entire human. This required working in primary iPSC-derived neurons. It required screening five million cells. As you remember, it's never just about the AI.
You also need to have the right data and the right experiments, and as an illustrative example, I can tell you that we can identify new gene pathways that cause a cell state that mirrors tau aggregation. We have also found many neurospecific perturbation effects, gene modules, gene functions, and associations with key disease phenotypes. From the AI side, these kinds of data mean that we can also train new multimodal models.
This is the second model that I will introduce today, and it is called MAP-LIB. The two modalities I showed you before with Nona were DNA and RNA. With MAP-LIB, we train on the RNA and the cell images from two perturbation screens, and then we translate between them. For example, MAP-LIB would learn the RNA profiles that correspond to different phenotypes that are captured by images.
Then for the next perturbation that we might do, we can just capture an image, which is a much more scaled experiment. It is easier to do. It is easier to do with small molecules. And the algorithm would then generate for us the RNA profile. What else can we do? Multimodal models of other types can essentially become CellO racle for us.
An example that's especially exciting for us as we turn from the target discovery to the compound side is a CellO racle that predicts if a virtual compound would phenocopy a known small molecule inhibitor or knockout, basically conducting a virtual phenotypic screen. This could help us make molecules for new targets or better molecules for validated targets. How would that work? We start by training an oracle model on prior small molecule phenotypic screening data.
We take the system that we're interested in, which might not be the same one in which we did that previous screen, and an anchor compound, say, a known small molecule drug or a genetic perturbation that has a desired phenotype for us. It would be a known drug if we're trying to improve that drug.
It would be a knockout if we're trying to discover a drug for a target that has never been drugged before. We treat the cells with them, and we measure the cell state, and we encode the cell state with the models that we trained. Next, we take a large-scale virtual library that can be in billions or tens of billions of compounds. We apply our computational oracle, and we make it predict the cell state.
We then choose as virtual hits those virtual compounds that yielded a matching phenotype to the anchor. To explain how that actually worked in practice, I'm going to show you an example of where we are now, so we trained a first CellOracle. Next, as the anchor, we used an established small molecule drug. I'm just not telling you which one it is.
We treated the cells with this drug, and we measured the phenotype of the cell. Now we want to use the oracle in order to score billions of other molecules for their ability to yield this cell phenotype. But this is the first time we're using the oracle, so we wanted to, first of all, test it. What did we do? We took the molecules in our standard screening library, and we had the oracle score each of them.
Then we compared that to the results of a conventional in vitro high-throughput screen we did, not a phenotypic screen, which we did years ago, this high-throughput screen for this target using our standard library of about a million, two million compounds against this drug target. From that screen, we followed up on a subset of molecules.
As expected, some of them were real hits, and some of them were not. Now, as you can see in the distribution that is shown on the left, the scores from the oracle were much higher for those molecules that were real hits. That is the curve in red. They were much higher for those molecules that were not followed up at all.
That's the curve in black, or the molecules that were followed up but were not ultimately deemed as hits, which are the ones in gray. The dashed line in yellow is the oracle score on the anchor small molecule drug for your reference. Finally, we took this oracle and we applied it in the true virtual setting where we can score many more molecules.
We got virtual heaps of multiple scaffolds, and that's what you see on the embedding on the right, which shows the virtual heaps for synthesis, and that work is actually ongoing. Okay. So this is a great place for me as we're getting into molecules to actually move to the next step for our lab in the loop, which is the lead identification and lead optimization stages. The basic principle of what I'm going to show you in this section is the same.
We're going to train models on data. We're going to use the models as oracle, just like the CellOracle I just showed you, or we can use them in generation mode as GenAI to predict properties or to make new molecules and design them. And then we test those new molecules in the lab, and those are then additional data also for our models.
This is active learning with an oracle. It's a very generalized concept, and it applies across all of molecular drug discovery. So I'm going to start by considering small molecule optimization. In this approach, data come from both high-throughput screening and from the optimization of prior molecules. The models we train on this prior data are used for optimization. The molecules are synthesized and they're tested in the lab, and then we repeat it.
This can be done for many chemical features, potency, safety, phys chem, clearance, and so on. For each of them, we have each of them can have its own oracle, and we can either work, you know, for one target with one or for a target on another, or do it sequentially, although ultimately we need to perform multi-objective optimization in order to balance across all of these properties.
A second component for small molecules is first principles generative AI, like our structure-based drug design with the denoising voxel grids, an algorithm called VoxBind, which I am showing it run on a cartoon example, on a non-portfolio illustrative example in this video, and so how does this approach work? This is a score-based generative model for 3D molecules, and it is conditioned on protein structures.
We represent the molecules as 3D atomic density grids, and then we leverage a 3D voxel denoising network for learning and for generation. In particular, we generate structure-conditioned molecules using a two-step procedure. In the first procedure, we sample noisy molecules from a conditional distribution using the learned score function. In the second step, we optimize clean molecules from the noisy sample with a single-step denoising.
Now, compared to current state of the art, this model is simpler to train, much faster to sample for, and it gives better results on many different benchmarks. It generates more diverse molecules. It has fewer steric clashes. It has higher binding affinity to the protein pockets. But the most important thing, of course, for us is that it actually made us better at drug discovery.
So for example, in a very recent portfolio program, it's not only that we transitioned that program from LI to LO with 25% less time by using this approach, but also the molecules that we achieved relied absolutely critically on the design of AI. We wouldn't have achieved them otherwise.
And while the movie you just saw was completely illustrative, the real program has a movie that looks exactly like that, just with our real target and our real molecules, and that is where our LO transition came from. We also tailor approaches for our macrocyclic peptides. As you know, macrocyclic peptide screens are iterative, so we run one screen, we get hits, and then we use those hits as the start of the next screen.
Our goal was to be better than the current method of choosing hits, which relies on a human expert to make the choice. With a new method called DeepF itness, we train an initial model on prior screens. It makes predictions, then we synthesize those molecules testing in the lab and iterate. That sounds very similar to what I just described to you for small molecules. But there are two critical insights here that are unique to this case. One comes from evolutionary theory.
We don't just rank by enrichment, but we actually rank by fitness. The second is the particular way in which we model off-target effects. And just like the VoxBind approach, here too, we apply DeepFitness, which makes us better. So we applied it here to a screen to find macrocycles that bind an important target in our portfolio.
And as you can see, it did a lot better than the human expert. In fact, it is the human experts themselves which are quoted on this slide. I would have never picked those best hits. But in the follow-on assay, the molecules from the AI, or chosen by DeepFitness, bind two orders of magnitude more strongly than the best ones that the human expert found.
What is especially important is that these hits are also small or on the smaller side of a macrocycle, and finding small macrocycles is crucial for identifying drug-like peptides that can move in our portfolio. Now, many of the same principles, but separate algorithms, apply in the optimization and the de novo generation of antibody therapeutics, the active learning loop, the oracles, etc.
On the oracle side, we now have the deployed oracles for expression, for binding, for immunogenicity, for PK clearance, but also for very specialized use cases like ATP and pH sensing, for example. For a given target, we can deploy, just like with a small molecule, we can deploy one oracle at a time or multiple oracles. We can do it sequentially, or we can do it in multi-objective optimization.
Now, our lab in the loop for antibody optimization continues to improve over many cycles of advances, and that's not just for the same algorithm over time, but also which algorithm and model we use and how we combine them during the optimization, and so in current iterations, the designs we get are basically invariably expressed.
They're at 100% expression. 80% of them are binders, and a good fraction binds with threefold or higher improved affinity compared to the seed antibody that was used to initiate the round, which was already in its own right a binder and usually with pretty decent affinity. This means that our engineers get a real boost from this process. Oops, I went one click too far, and I want to highlight for you just one of these oracles.
The specific one I'm highlighting is called DIAB, and we use it for affinity improvements, the process I just described on the previous slide. The problem that DIAB is set to solve is the challenge of the very large space of possibility of combinations of mutations that we can make in the variable regions during optimization. When we start the process with DIAB, we usually have already had experimental affinity training set, which was based on prior iterations and in particular on testing up to, say, 100 single and double mutants in a short variable region.
We sometimes find some binding improvements in this experimental way and might be done, but more often than not, we would want to do more. What we would really want to try out are many, many mutations simultaneously.
The problem is that testing all possible three mutations in a region would mean about a million constructs, and testing all four mutations, all possible combinations of four mutations in a region would mean about 10 to the 8, none of which we can do experimentally, nor can our engineers actually imagine all these options in their head. In fact, even computationally, some of these numbers become a problem if you're going to enumerate them one by one.
And so what do we do about this? Here comes our DIAB-guided machine learning. This particular oracle predicts affinity differences instead of just affinity, affinity differences between two similar sequences. And when we combine it with a search algorithm, like a genetic algorithm, it can propose novel sequences that have higher affinities than the started sequences.
Going a bit deeper, the idea behind the algorithm is to utilize a protein language model embedding, actually two of them, to represent the difference between two protein sequences and then estimate the difference between their predicted properties. We trained this model on a combination of publicly available antibody sequences with fine-tuning on internal affinity data, and it has been able to generate binders with 10- to 100-fold improved affinity compared to the starting seed for multiple targets in our portfolio programs, even exceeding in some cases yeast display.
Okay, so these are optimizations from an existing antibody, and in parallel with de novo antibody design, like the movie I'm running here, we can use specific models within the very broad framework of diffusion models to generate realistic samples of therapeutic-like antibodies, which can then be scored by the relevant oracles for design.
I'm not going to belabor an algorithm example on this one today. Oops. Let me move to the next slide. Okay, so to close this section, I want to bring you back to the first model that I mentioned. That's Nona, the multimodal DNA and expression model.
Now, I described it to you in the context of a task that has to do with target discovery and variants in regulatory regions from human genetics, but as you know, part of the definition of a true foundation model is that it can be applied to many tasks, and it is not just trained to perform one particular task in a narrow way, and so we ended up reusing Nona in multiple settings, and one of them is in gene therapy programs. The specific challenge here was twofold.
If we deliver with a viral vector, we have a limitation on the size of the cargo. And so the team wanted to make the promoter as small as possible. But at the same time, we want the cargo to be exquisitely specifically expressed in a particular cell type or particular cell types. And so we also want this promoter to behave like a good enhancer that can drive very specific cell type expression. And that usually takes some real estate.
And so the team put Nona to work in its generative capacity in order to reduce the length and generate cell type specific expression, and it handily achieved this: a shorter promoter and higher specificity. Okay. So now let me turn to our last part into the clinical lens and our patients and the opportunities really to think about a clinic in the loop.
One key component is getting to the right indication. Last year, I actually told you in a similar event about the first version of our foundation model for cells, which was called scimilarity and encompasses hundreds of millions of cell profiles. During this year, we actually expanded scimilarity by pairing it with human text so that scimilarity can learn how scientists describe specific RNA profiles. One of the application areas is indication selection or expansion.
How does that work? We start with a compound with some indication and the expression profile it generates in cells. Using scimilarity, we search our large model across internal and external data for cells with this profile and the indications that they came from. Last year, I told you about one expansion like this for vixarelimab in IBD. Now, actually also augmented by text, we use it in multiple other programs in the portfolio.
A second key approach for us is to bring AI to enhance both the cost and speed of clinical trials, but also really critical decision-making. And a central opportunity here is our clinic in the loop for imaging, which I'm going to illustrate in our major therapeutic area of ophthalmology. So in ophthalmology, we have multiple forms of non-invasive imaging, and those are a really critical component of our patients are monitored and our patients are assessed.
One of them is called fundus autofluorescence or FAF. This is what you see on the left, where we can assess these lesions in two dimensions. The example here is in geographic atrophy. Another one is called optical coherence tomography or OCT. We see this on the right. Here we have high-resolution cross-sectional images in the eye, in this case in nAMD or in DME. All of these are prime territory for AI.
Everyone knows that images are a big deal for AI. Now, in the case of FAF, both humans and machines can segment these kinds of images, but once we train the AI on a ground truth, it will do this better, faster, and cheaper than a single grader. In the case of OCT, AI is even more appealing to us. Layer segmentation and the kind of quantification we need to do for fluid requires a lot of training. But even if you are a super-duper trained person, it is infeasible to do this at scale in 3D by humans because for a single 3D image, it's actually a series of many, many, many, many, many, many sections.
And so AI that is trained on the ground truth for a subset of the 2D scans would not just give a scale and granularity, it would give a scale and granularity that are not possible otherwise. And that opens up possibilities. Well, how does this play out in our work? In the GA case, the AI graded more than 10,000 FAF images from historical trial data. It saved millions of CHF that it would have cost to do manually, and it accelerated us more than a year to design a new trial for a different population.
For the NAMD DME case with OCT, the AI graded more than 180,000 images. That's why humans can't do it, from the Vabysmo trials. This helped support treatment differentiation, and the features helped us characterize populations that have remaining unmet need and are crucial for our future programs.
This is all great, but now let me raise the bar further, and so let's focus on an even more challenging task. Can we build a model that would robustly predict the future growth of the lesion from the baseline image? And it turns out that we can.
We trained an end-to-end model like that to predict the future growth of GA lesions and actually the growth rate of GA lesions, and we validated it then on 2,000 images from our previous geographic atrophy trials, and this model explained 48% of the variability compared to less than 30% that was explained by a traditional model that was built based on features that were derived by human experts. That's a 60% improvement. But what do we do with this knowledge? As you may know, the geographic atrophy growth rate of the lesion is actually the primary endpoint for GA clinical trials.
The efficacy there is assessed by the mean difference in lesion growth between two treatment arms, the control arm and the treatment arm. And as always, if there are imbalances in the arm, for example, if the patients in the control arm would actually be ones that progress faster just on their own, those imbalances can mislead us. If the control arm progresses faster just because its baseline was destined to progress faster, the trial could read positive, but in fact, the molecule is not efficacious.
And so this is a pretty big deal for us. The AI model that we have trained can predict the growth rate variation, and so we can use those predictions in covariate adjustment. And as a result, we would have a better treatment effect estimate or better removal of any imbalances.
To show you how this plays out, I'm going to use a retrospective example. The example comes from lampalizumab. This is a molecule that we had in our portfolio in the past. We're looking here at the results of two trials. On the left is the MAHALO phase II trial, where we had a seemingly positive effect, and the molecule was advanced to phase III.
On the right is the unfortunate outcome of the phase III trial. It did not confirm the phase II result. There was actually a lack of efficacy. The imaging data from this trial is actually some of the imaging data I showed you that has enabled us to really advance in AI for geographic atrophy.
And so now we went back and we used AI models for covariate adjustment, something that simply could not have been done at the original time of this trial. And what we see is that the false positive effect in the phase II goes away. Had we had it at the time, we would not have advanced this program. This is what I mean by having AI help, not just with efficiency.
It's not just saving money. It's not just saving time, but with our most critical decisions. Another way of thinking about the kind of prediction of effect estimates that the AI allowed us to do here is that it is as though we ran a study that was more than twice as large in terms of its power around efficacy estimates in this specific example.
Now, we can also use AI-based approaches, not just in ophthalmology and not just with static images. We can also use them in movies. So, for example, in our end-to-end disease of IBD, we use AI on colonoscopy videos to score spatial disease severity to improve the assessment in ulcerative colitis. The AI helps us here with multiple tasks, including determining where we are actually in the colon as the scope video pulls through. There is much more that can be done because in clinical trials, it's not just that we have clinical trial results.
We also have real-world data. We have the electronic health records, and so much can be done with them all the way to digital twins that are based on GenAI. And those become then virtual representations of real patients to explore, predict, and simulate their health trajectories and to help us with clinical trial design.
And so this example actually comes from our colleagues in Roche in Basel, who recently reported the development of DT-GPT. DT stands for digital twin. And here they combined biomedical large language models that were fine-tuned on electronic health records in order to perform better on tasks like forecasting clinical variables and exploring different kinds of predictions of the digital twin. The team, in its assessment of this model, used both long-term non-small cell lung cancer data from the U.S. as well as short-term ICU data.
And they showed that DT-GPT actually surpasses other state-of-the-art machine learning approaches on many things: on patient trajectory forecasting, on being able to preserve the cross-correlation of clinical variables, and also on handling the missingness and noise in data that is really inherent, especially with real-world data and the EHR.
And so, in addition to that, and that's very cool, it actually managed to provide insights into the rationale behind the forecast. The model could explain itself, and it could do zero-shot forecasting on variables that were not used during the fine-tuning.
So now we can start imagining how these LLM-based digital twins, including with LLMs that are trained on the most relevant internal data, could help us with clinical trial simulations, with treatment selection, with adverse event mitigation, and overall enhancing our clinical trials and the success that we want to deliver for patients. And so with this, I will close. I will thank you very much, and John and I will be happy to take any questions. I will stop sharing my screen.
Thanks a lot, Avi, for this very insightful, exciting presentation. Waiting here for the hands to go up. I hope we don't have overwhelm. We have a first person who wants to ask a question. Can you please clearly state your name and the company or university you're working for?
Hello?
Hello, there's one person in the call who wants to raise a question. I have opened the line. You can please ask your question and please give your name.
Doesn't seem to work.
Hello? Hello?
I've opened the line. Could you please ask your question? Doesn't look like. Okay. Yeah, here we have now the first question from Steve Scala from Cowen. Steve, please.
Thank you so much for a really interesting presentation. I'm just wondering how Roche's efforts in this area, do you believe, compare to those, say, among competitors or in academia or government? Are Roche's efforts unique in any way, or are they more uniform throughout the industry?
So I can never exactly know what is happening in other companies. I think I can comment quite reliably on the academia side and what the government is doing, but in other industry partners, it's always harder for us to tell.
So I'll caveat that. I believe that in several of these areas, we are absolutely at the very forefront of the state of the art, and I think this is reflected in things like the kinds of papers we put out, the publications we have, and the performance that we see internally, including in our ability to really do it internally rather than rely only on partnerships for it. The reason I think that is really important is that the models really rely on training with a lot of data, and that quality and quantity of data is hard to share whole cloth with external partners.
I will also say that as a result of it, anything that happens in the learning loop doesn't just provide for us a target or a molecule. It actually provides to us also the better algorithm for all targets and all molecules. So in that area, I think we were a very early investor, and we built an extremely strong team.
They have a lot of external visibility in the best places, and that is the way in which I can measure that I do think that they are at the cutting edge. When I compare to academic efforts, what I described for molecules is very, very hard to do in an academic setting. I think some of it is hard, and some of it is on the verge of impossible.
And the reason for that is that the training data don't exist in the scale and scope that they need to exist because this work is almost invariably done by companies. In small companies, they don't exist because a small company doesn't have decades of history for its data, and the ability to do all the iterative work does not exist. The amount of synthesis you have to do, the number of iterations you need to do on the synthesis, and so on. And if you look at where the literature is on that, which is in conferences and talks, I think it's pretty illustrative of that. You see a lot of work happens where the data are.
On things that you can do, for example, on protein structure, you can do a lot because a lot of protein structure is in the public domain, but antibody structures are not like that. Antibodies are a very unique creature. They're not similar to others. We're interested in the most variable regions where evolutionary information is not as helpful and where you need to have complex data and where you need to know the epitope.
That type of richness really is hard to mimic externally and requires a substantial amount of investment. When I turn to the biological side, again, I think we were not only early adopters. A lot of the pioneers of these methods experimentally and computationally work in Genentech. By that, I include the two people who are on this call.
That has given us an edge. But everyone in biology works in a more open way because data tend to abound more publicly, and the iteration happens in a more open way. Where our competitive advantage lies is in having patient data from clinical trials. I alluded to that when I described the IBD case.
That is invaluable and very hard to achieve with detailed clinical information and very high-resolution data, and in the commitment to combine it together with drug discovery, like the work I described that we do with large-scale screens to build the CellO racles. Again, an area that is very, very hard to replicate in academia and mostly does not exist there.
Steve, did we answer your question? Do you have any follow-up questions?
Did that address everything? I didn't get to the clinical side. I think in the clinical side, it's pretty obvious. Again, having these highly detailed images with clear clinical annotation at that scale that comes from a high-resolution clinical trial and the ability to deploy it in a clinical trial, that is something that I think large pharmaceutical companies are all committed to and probably work on intensively, but you have to stay at the very cutting edge.
Thank you so much. Okay. And the next question goes to Emmanuel Papadakis from Deutsche Bank.
Thank you for taking the question, and thank you for the very helpful presentation. I've got several questions. Maybe I'll start with a couple. Perhaps just thinking about R&D spend, what percentage of the R&D budget is being spent on AI tools such as those you've described today? And does that drop at some point and presumably reach a peak in terms of investment in these capabilities? Does it then level out?
Does it grow in line with the overall R&D budget over time, or does it actually drop? That's question number one. Question number two is, do all projects in gRED now by necessity or not, rather than necessity, by requirement embed these components? Do you force all teams to use them? And how many clinical candidates in the gRED pipeline have used components of the approach you've described today so far?
So I will answer question one. John might answer a portion of question two, and John will answer question two just so that we divide and conquer. There are two components to what I described that impact the gRED budget, the gRED research budget, and thank you for actually separating it from the development budget. One major component is the people component of the AI together with our compute needs, etc., etc.
And the second major component is actually the data generation and iteration, which is actually sizable and very needed. For the first one, it is easier for me to do it off the top of my head with precision, which is that the size of our computational organization that John leads is almost the size of our drug discovery organization. And I think that tells you a lot on the size of the investment, so it's substantial.
On the second one, it is a little harder for me to do it off the top of my head, but a sizable fraction of the increase in our R&D budget in gRED over the past four years went actually to fund large-scale data generation capabilities, including the people, the instruments, the infrastructure, and the experiments themselves, such that we collect data not just for programs now.
We collect data for foundation models now. That has been a big shift in how we do things. For your question on how it will evolve with time, I will say that the data collection piece, I venture to say no one in the world knows right now how it will shape up. It depends on the generalization ability of these models, and we do not understand it scientifically well enough for every single case.
There are cases which might have reached their plateau. For example, if you think about native states of cells in the human body, something that I have happened to do a lot, I think we're pretty close to that plateau. Whereas for other things, like the response of cells to different small molecule perturbations, I am certain we are not.
So there is quite a gap between those two things in terms of how much more would you ever be continuing to collect data because over time, you will collect less data, and the algorithm will generate more. From the algorithmic side, I think we still have many, many problems for AI to solve after the ones that I've described, especially as you move to the higher order levels of reasoning, like counterfactual.
We're barely scratching the surface in AI in general, not just in AI in pharma or in AI in biology of these kinds of problems, and they are material for the biggest successes that one can imagine. So there's still going to be work to be done. For the portfolio, I will tell you that I showed you advanced examples, but these are not yet clinical molecules.
The things that we have in our portfolio that you can say AI did a direct impact on is our vixarelimab program, which we would have not done without large-scale data and machine learning, and our cancer vaccine, which is designed actually in large part by state-of-the-art AI algorithm. And I'll turn to John to talk about how this plays out daily in gRED. Yeah. So thanks, Avi.
So I think that in the molecule design part, this is really very well embedded both in the large molecule and in the small molecu le side. We're deploying lab in the loop routinely. And as we've mentioned, it is leading to candidates that are beginning to enter the early stages of our pipeline, and that's really exciting to see that come to fruition.
Scientifically, that's probably the most mature area. You saw the Nobel Prizes that were won this year in this space. And so we've been able to really advance that. On the biology side, we are again making really great strides in the target identification. Some of that is on the methodological side, but there's also that data generation part, and it's tempting to forget that, but the investment in the data generation is tremendously important, but also in bridging the disciplines.
So for example, in the oncology space, we recently brought in a fantastic scientist, Anwesha Dey, into a leadership role that really bridges the wet lab and the computational to ensure that we are deploying and generating the data at the scale we need and on the right questions to build those loops effectively. So we're beginning to see those new targets begin to come out. We're beginning to really see the impact in the biology space.
It is becoming just increasingly part of life, everyday life for the science that we do, that this is going to be part of the future of GRED and, I believe, part of the future of the industry, that this need to couple data models in a really cyclical way that's going to improve target identification, is going to improve molecule design, is going to improve our discovery of better biomarkers, is just becoming part of life.
That's been one of the big changes, I think, over the last couple of years where it was sort of, it wasn't academic, but it was kind of people were a little curious about it. Now it's just becoming part of the furniture, and you begin to see those outputs, and as people see that, they're increasingly excited about deploying it across the whole range of activities that we do.
One addition on your portfolio question. I believe the large molecule with AI inside is just making its way to what we call a D ev Go, right, which is a clinical candidate has been chosen. That is, I think, from the large molecule side, the most advanced.
And by that, look, you had computation and AI in molecules for many, many years, but it was old generation machine learning, a random forest or other optimization approaches that people used, and so on. What I refer to is the real lab in the loop. So it's a different style, and that is all I count when I give these examples.
Very helpful. Thank you. Okay. Are there any additional questions? Emmanuel, yes, please go on.
Sorry, me again. I'll take one more if I'm permitted. If I was going to slightly play devil's advocate, I would say it still appears at the stage where we're providing tools that complement the human role in the drug discovery process rather than replace it, right? So you're providing teams with additional tools to do their job in a better way.
But surely there's also the risk that they've become a source of distraction and inefficiency if in the end, they don't help them to do their job better. They're just an extra thing they need to work with and a tool they need to interrogate or employ in their process. So how do you manage that risk that this is not actually diluting efficiency for some teams? Presumably, in every case, it doesn't necessarily work the ideal way you would hope or intend.
So I'll start, and then I'll turn to John because I think it's a material question. First of all, it depends on the assumption of who's better, the AI or the person. And one of the things I showed here, and I've showed in other examples in the past, is that we have a non-trivial set of examples where we can quantify, right? We compete them essentially on the same data, and the AI comes out on top, and that they tend to give different answers.
And that suggests that you actually want both if your focus is on PTS. If your focus is kind of on the narrowest definition of efficiency, like how many resources do I put on something at once, and you consider the AI expensive, then you do one or you do the other, but you don't do both.
But if your focus is on the PTS, then you do both, and you get an extra edge from that. It's different. We have seen this with agents a little bit in some appropriate head-to-head comparisons that they actually the agent is better than our average human and slightly below the best human, but the best human plus the agent are better than the human.
That's the examples that really motivate us. And honestly, once developed, the cost of the inference is not that high. And as a result, it's actually, why not use if you can be better than what you are today? From an efficiency point of view, from timing, one of the critical things is actually building an R&D process for which this is not a distraction. It's part an d parcel of how you work.
Because initially, the main reason these things are distractions is because people are not exactly sure how to interact with them, because teams ask a lot of questions, because they need meetings in order to coordinate and so on. But as John said, once it becomes part of the furniture, it's actually a speeding factor, not a delaying factor.
At the same time, everything needs benchmarks. Things need to be tested. The fact that they're promising doesn't actually mean that they deliver. So you do need ways, just like the ones I described to you, like competing against humans, doing other things in order to really know that something is delivering what you thought, because if it is not, you should retire it and look for something else. John, I'll turn to you.
I'm going to unmute switch. There we go. Yeah, I'd agree with everything you said, Avi. I think that ensuring that it is complementary is very important. I think that as people see the applications, there's a lot about how people see something and when they see it really work. We're scientists. We kind of learn through that, and then you want to be able to deploy it more.
And as we begin to see more and more of those illustrations where it really does have a very positive impact, you see that people are really very enthused about adopting it, and that adoption goes up, and they learn how to use it effectively. And over the last years that we've mentioned with the development of agents that are really very powerful, it also democratizes the tools, and so it makes it easier. The bar to entry is lower.
You can engage with it just in real text in the same way that you do when you're organizing your holiday itinerary using some of these approaches. And you can begin to then really deploy it without that very large additional bar. And that allows people to really think about how they're going to d eploy it in an effective and real-time way.
And that's something that I think we're also investing in because we believe that that's going to be very powerful, not only in developing the data and the technology, but democratizing those tools and making them accessible so that they really aren't a burden on people's everyday work, but just part of what they do, and they're able to deploy them r eally straightforwardly.
And beginning to see that come in and see some of the real enthusiasm that you see from colleagues about how easy it is to work using these agents just in this tech framework, that's really exciting because it does open this up to everybody. And the additional burden and the distractions that you're worried about with people being sidetracked, they begin to dissipate.
So the tools get better, the approaches get better, the ability to deploy them gets better. And as you begin to then see those examples that we've talked about where they really have a positive impact upon the pipeline, people will just and are deploying them throughout. You really do see that change over the last couple of years on how people are taking these approaches.
The two examples that I showed, just to give you a real flavor for them, even though I couldn't say what the target is and I couldn't show the actual molecule, which obviously as a scientist is slightly frustrating. The example I showed with the VoxBind, the generative model, that was applied in a very, very high-value program for us.
It cut the time by 25%, which is awesome, but what is also really awesome, there wouldn't have been a molecule otherwise. There just wouldn't have. The thing the AI suggested to do, no one came up with. And so the medicinal chemists were very happy. That motivates them to do the next molecule naturally, but it's a critical program for us, and now we have something, and we wouldn't have something otherwise.
The macrocycle screen that I showed, the screen itself yielded zilch in the original way in which the heats were chosen, so these are things that are like nothing or something and substantial acceleration. The example I showed with the CellOracle, even though it's still work in progress, right? The molecules are synthesized, but they haven't been tested yet, so I can't tell you how well they actually work.
We didn't have a way to do virtual screens for phenotypes, for phenotypic screens. We just had nothing, so if you don't know the target, you can't do a high-throughput screen in the standard way, and doing a high-throughput screen is a very expensive and large-scale endeavor, being able to do it virtually when you're still in target assessment and all of a sudden have leads, and you do it.
Once you have the system, you can apply it now again and again and again and again. So those are the things that really motivate our people to engage with it, but also that motivate us to say this really changes how we do stuff. It's not just one more distraction for a team that has a big menu of options. Very good. There are many more questions, so we should stop on this one.
Yeah. Thank you very much. Very helpful.
Then a follow-on question from Steve Scala from Cowen.
Steve, please. Thanks for the follow-up. You focused mainly on positive outcomes, but are there examples where AI generated the wrong answer? In retrospect, why did that occur, and how was it detected?
Oh, actually, when I showed the graph, you could go later on and look at the slides of the iterations of the lab in the loop. You can see the places where it totally crashed and burned. It wasn't like expressing. Nothing was binding.
Or they were expressing and binding, but the affinity was like tanking. We have graphs like that for everything. The reason we track them, we have a dashboard that tracks every iteration is because it actually fails before it succeeds. And that's in the molecular sense. I'll get to the biology in a second. The reason it fails is that it can be varied, but often it can be algorithmic choices that are not set up correctly. It can be going too much off the deep end in being generative all the way into the hallucination domain. It really depends, but it happens a lot, and it is valuable.
That's the part that I think is very, very hard initially for experimentalists to accept how useful the failures are, even though when you run an experiment as an experimentalist, you learn from the failures all the time. The human brain does the same.
But those failures are really important, as is the need to synthesize, make, do experiments for things that the model is uncertain about because the model wants to improve with certainty. It doesn't just want to make you a molecule. And that's part of the process of understanding that. I'll say in a moment something about the biology, but I'll turn again to John for a second for his perspective. He sees it like day in and day out.
Yeah. So it is hard when you predict something and it's negative, but as Avi said, it's actually really powerful and important for the model. Learning what doesn't work allows it to then generate better predictions in the next iteration, and you also see that on the graphs that Avi showed.
Yes, you do have those cycles where things didn't work, but then that allowed the model to explore a better area of the space, and then that allows it to then make better predictions, so this negative data is something that's very important. It's something that we really treasure as well, and having that internally and that access to that large amount of data, both positive and negative, is really critical. On the biology side, I think that I'll have to have even a second, but the generalizability is sometimes rather challenging.
That's just because in the biology space, the volume of data that we have compared to on the molecule space is much less, and it tends to be from generalizing to broader populations or to unknown perturbations is harder just because we have less training data. So there, I think the challenge is how generalizable are the models. So how could we apply them, say, to an individual with a different genetic background?
And do we have enough data to train the models well for that? We're generating that, and we have more of it. But I think when we look at the effectiveness of the different approaches, some of it is around the generalizability of the models that we have and how well they are able to be applied in other contexts.
But as we generate more data through those partnerships with Recursion that Avi talked about, through our own internal efforts, through other public sources, the models will get better and better and more generalizable, and that will then allow them to be applied in a variety of different contexts. But at least from a scientific perspective, that generalizability at the moment in the biology space is hard because the amount of data that we have relative to the chemical space is just that much more limited when it comes to having a broader range of training data. But I'll pass back to Avi for her perspective on that.
I'll start actually with something, Steve, I think you asked earlier about how we compare to others. And I pointed out that there is a limit to what you can do in collaborations on some of these problems. One of the limits is the need for these many iterations through failure until you get to really amazing success and that those sometimes don't stick, right?
You try with an external partner, you give them the data, they're supposed to deliver some molecule, they don't succeed in the first round, they don't succeed in the second round, they don't succeed in the third round, they don't succeed in the fourth round, and maybe it's going to be amazing by the 15th, but by the 15th, the partnership will be closed because the team will look at it and they will say that they can't deliver, so some of it requires actually sticking by it for several years, and then when it works, it starts working for a lot of things, and it becomes a lot more impactful as a result.
But those are learnings that you have on the inside. The second piece on the biology, John covered really, really well the generalization problem. I will highlight a different one and the way that we're thinking about it now, which is the volume. These experiments in biology generate massive amounts of results, all of which could be completely valid.
It's like endless results. And the challenge becomes not learning a model from it and not delivering it to the not even its accuracy, but we don't a human doesn't know what to do with that much. It's almost like too much. And then it's hard to reason through it. And there's a huge literature to reconcile together with the results of these massive experiments after the AI, after the model, after everything.
So that is where the agents actually become so impactful because A, they liberate a much larger number of people to act as though they're computational biologists. They don't need to know how to code and how to run and which tools are out there exactly and so on.
And B, because you can ask the agent to reason through this as well, and you can develop agents and models that really try to do next-generation reasoning on these results to say, "Oh, when I do this, that happened because this might be the chain of events that led there," and so on. These are really time-consuming and difficult things, and there's no way, I mean, for a human to do it even in a lifetime, let alone in the speed that we actually needed for drug discovery.
Thank you. Steve, did this help?
Yes. Thanks very much. Thanks, Steve.
Can we move on? We have another question from Anurag Dhanwantri from Brown Brothers Harriman.
Can you guys hear me okay?
Yes, we can hear you.
Okay. Thank you so much for doing the call. I actually have more back-end technology-related questions. Given all the data privacy and regulatory issues, are you able to train your models on the public cloud, or do you have to do it in the private cloud? Just curious how that balance works out. And the second question is on the sheer compute power you need. Are GPUs becoming a bottleneck for you? Is the pace of sheer compute power that is getting delivered by everybody now a bottleneck, or are you able to get the resources that you need? Thank you.
Yeah, and I think you can get started on this one. Aren't you surprised?
Yeah, I was going to. I wondered if you would pass that one. Let me take the second one on the compute power first, and then onto the privacy because it's a little bit more nuanced perhaps there. On the compute power, the models that we're talking about, especially the training aspects, require very high-end GPUs.
The type of infrastructure you need for inference and some of the other fine-tuning might be a bit less, but for the training of those foundation models, yes, that's really critically important that we have access to the most sophisticated computing because it just makes the ability to train those models in a reasonable period of time just much better. That's why we have specific partnerships externally.
So working with our colleagues in AWS, but also with NVIDIA, to have access to that sophisticated compute so that we are able, as needed for those really high-end models, to have access to the necessary GPUs and the number of GPUs that we need in order to train them in real time.
I think as we move forward, this is an area that we're always keeping an eye on because it's moving very fast, as you know. And so we want to be able to get access to that compute as we need it to train these foundation models in a reasonable period of time. And then maybe I'll pass to Avi for the privacy aspect and the public-private cloud aspect of the question.
So as usual, in this domain, data, algorithms, compute. You need all three. I think on algorithms, we described at great length the investment, the work we've done, and so on. On data, I talked a lot. Compute is absolutely part of the equation. As John said, we've actually invested in partnerships in this area. This is one where partnership is critical, and we were early in it and very invested in it and very grateful for our partners. Part of it is in having access, but part of it is also in doing your algorithms and your training and inferences efficiently.
So you actually reduce your compute needs by being very efficient in how you do computer science. And that is part, actually, of our partnership with NVIDIA is to achieve these efficiencies. That also reduces its cost, which, as you know, we don't want it to actually become a more expensive way of running experiments.
Running them virtually has to be cheaper than running them in a lab. Otherwise, it doesn't make any sense. For the privacy point of view, Roche has an extremely robust approach to how to handle privacy, how to access data, how to maintain data, how to maintain it with the appropriate security, how to address all the ethics and all the rules and the regulation, and very early on, when we started this journey with AI, it is actually a very close partnership with our legal and IT colleagues so that everything that we do actually addresses this, and this includes what goes private, what goes public, cloud, and so on.
Even in public cloud, you are usually in a very closed garden of your own, and so even there, there's a lot that you can do very safely. But these are things that are done all the time, continuously, in order to manage any aspect appropriately from a legal and ethics perspective.
Thank you so much. Did we answer all the questions? Okay. Then I think we are at the end of today's call. I would like to thank you.
One more thing. Sorry. We have one hit.
Maybe there's one more question. Yeah. This goes then to Kelly Close.
Oh, thank you very much. And this has just been absolutely a phenomenal hour and a half of learning for all of us. So I really salute you. Just to end on a broad question, I was wondering if you have thoughts about where this is kind of more helpful, all o f this work, as far as therapeutic areas go.
I loved how you were really careful not to tell us necessarily which area certain molecules were in and so forth, but if you had any thoughts broadly, or if i t's just great for everything, whether you're talking oncology, endocrinology, maybe even many areas that are devices, etc. Thank you very much.
I will answer very quickly all of the above. Actually, last year, I think I gave more examples from oncology on the biology side. It's more we chose some examples that would fit in a limited amount of time, but all therapeutic areas would be in play, and I would say all therapeutic areas and all therapeutic modalities would be relevant for this.
None more than others.
No, I think none of them wherever there is unmet no. Wherever there is unmet need, and for us, whatever our strategy is focused on is, of course, what our pharma strategy is focused on. But wherever there is unmet need, it's the same set of problems. You have a biological system you have to figure out. You need to understand the targets, the pathways, the cells.
You have therapeutics that you have to design. You have to deploy them in clinical trials. So in fact, this generalizes. In terms of the modeling, as John said, he already highlighted that the modeling in some domains generalizes more than in others, and that really depends. There isn't a single answer. Yeah.
Very good.
Thank you.
You're welcome. Maybe we just wait a second. If there is no final question, then we would close the Q&A. So I would like to thank our experts for all the work and the preparation and the insightful Q&A we had. I would also like a couple of people who were working on preparing the event from the IR side. It's Jon Meyer , who had the overall management.
It's Anna Hubelovska, who worked on Avi's deck. And then I think from the back office, Melanie Wolf and Eva Losert. If there are any remaining questions, then please reach out to the IR team. We are happy to assist and follow up. And with that, I wish you a go od day. Have a good day. Bye-bye.
Goodbye.