Good afternoon, everyone in here, and also good afternoon to everyone out there. I guess for some of you it's, it's good morning; some of you it's very early morning, and some of you it's even evening. I can just say I'm super pleased with the very broad attendance that we're having for our event today. So welcome to the Evaxion R&D Day. I have been looking very much forward to this. Today we will be taking the opportunity of going into details with our AI-Immunology platform. I'm Christian Kanstrup, I'm the CEO of Evaxion, and I also just want to introduce my co-host for the day, our co-founder and Chief AI Officer, Andreas Mattsson, who will be helping me taking us through the day.
Yes, and I'm truly excited. I mean, I've worked for this moment for 15 years, since I started in 2008 with Niels Iversen Møller . And here we are, so I really hope today will give us a lot of good interactions with you guys, Q&As, and also I hope, really hope that you get inspired of our AI-powered research.
I can only say Andreas said early on today.
See, I've been looking forward to this for 15 years. Finally I get to talk about all the things we can do, so now is the time.
Definitely.
And how are we going to do it? What we have been doing is we have been putting together what we believe is a super exciting agenda for the day. It's structured in four sessions: first, an introductory session, covering a little bit about Evaxion and the platform, but also, very importantly, about a key building block of our AI-Immunology platform, which goes across all the different parts of the platform. So super important introduction here as well. Then we will be zooming in on the infectious disease side of the business. After that, we will be talking about the personalized cancer vaccines. And then the final, session four, is going to be about some truly novel precision cancer concepts, which we are looking very much forward to be discussing with you.
This, this is not only going to be a one-way presentation. This is hopefully, as you say, going to be a super interactive session. So we will have Q&A sessions, and the way that's structured is the first 1 is at the very end of session 1. But after that, in each of the following sessions, after each individual presentation, we will be having a Q&A session. And the way it works is simple for you guys in here. You just raise your hand and ask a question. All of you guys out there, please just post a question in the chat, and we will pick them up here. So a lot of interaction, hopefully. It's a good way of getting to ask the questions you have on, on your mind, and we are looking forward to answering those. So that's how we're going to do it, right?
Just one thing before we get going, this is this slide. We will be talking about the future, and you all know—don't leave—when we talk about the future, it's uncertain. So let's keep that in mind when we go through the presentations of the day. And the reason for you not leaving, right?
Mm-hmm.
You want to share a little bit about what you are hoping beyond what you already said that we're going to get out of the day before we dig into the introductory session.
Yeah, exactly. So I really hope that you get the perspective of this is hard work, and also that we started so long time ago, that we now have and are ahead of the game. And we really now we're really benefiting from it, and this is really crystallizing in something beautiful. And, of course, it's getting saving life and improving life with AI-powered immunotherapies. That's what we're doing of AI-powered vaccines also.
Yeah.
That's a good expectation for the day. So let's do that. Then let me just start out. I mean, most of you know Evaxion, so I'm not going to spend a whole lot of time talking about Evaxion, but I'm just going to spend a little bit of time setting the scene, right? So who are we? Who is Evaxion? We are a pioneering TechBio company. Important: it's not biotech, it's TechBio. It might sound like semantics, but it is an important difference. And we have a validated and leading AI platform, AI-Immunology, which is what you will be hearing about today. And this is for fast and effective vaccine target discovery, design, and development. And with AI-Immunology, we can develop groundbreaking, both personalized and precision vaccines for cancer and infectious diseases.
One interesting thing about Evaxion, which is also important when we discuss today, is when Andreas and Niels, 15 years ago founded Evaxion, it was actually as an AI company, as an AI company with the objective of decoding the human immune system. It's also fair to say at that point in time people were looking at you guys smiling a little bit, thinking, "Yeah, that's fun. You can't do that. That's never going to work." Of course, fast forward to today, everybody is claiming they can do AI-based target discovery. We have the difference. We have been doing this for 15 years, constantly developing and refining our model and our capabilities. So I guess it's fair to say it was worse. People were smiling and laughing a little bit about you, but now we are in a good position. So a pioneering TechBio.
Also important is why are we here? We are here for saving lives and improving lives with AI-Immunology. I have to say that is what gets me out of bed every morning, thinking, "Today I will, together with all my colleagues, be working on saving and improving lives." If anybody doubts that there's a need for us, look at the numbers here: 10 million deaths a year due to cancer. That is a huge number that we need to improve. Also, 8 million deaths a year due to infectious diseases. Huge number. We need to improve that. We are here for saving and improving lives with AI-Immunology. How do we do it? What's our strategy? We have what we call a three-pronged business model based upon AI-Immunology, focusing on realizing value via a multi-partner approach.
So the way you should think about it is the core of Evaxion, that's AI immunology. That is what we will be talking about the rest of the day today. And AI immunology, that is about the target discovery, design, and development vaccines within cancer and infectious diseases. What's also important when we look at AI immunology is we can design a new target in just 24 hours. If you think about how long it takes to discover a new target with traditional methods, doing it in 24 hours, it's a significant and very strong value proposition towards partners. So discovery, design, and development new vaccines within cancer and infectious diseases in just 24 hours. What's also important, and that's another strong value proposition towards partners, is with delivery modality agnostic, meaning whether a partner prefers peptide-based delivery, DNA, mRNA, it doesn't matter.
It doesn't matter because the target that we design and discover with our AI-Immunology platform, that is delivery modality agnostic. And finally, and what you also hear about today, and which is super important and something we haven't really talked about a lot, is we have a modular architecture. Our AI-Immunology platform is built up of building blocks which can be combined in many different ways to give the desired outcome. So it's not like we're doing a lot of different things. We're actually having some very unique building blocks which we put together in distinct ways to generate certain outcomes. What does that mean? That means, of course, we have a high degree of scalability. It also means that we have a flexible model, which is important to those partners. So the core AI-Immunology, core of our strategy, how do we generate value?
That we do via the three prongs, as we call it: its targets, its pipeline, and its responders. Targets, that's, multi-partner approach to either single or multiple target vaccine, or vaccine target discovery, design, and development agreements. Teaming up with pharma, developing novel vaccines. An example here, that is the collaboration we just announced a couple of weeks ago with MSD, where we have partnered up to develop a novel vaccine against a bacterial infectious disease where no vaccine is available today. That's what we are doing here, teaming up with pharma, addressing significant unmet needs. Second element, that's pipeline. It's, progressing high-value compounds into either clinical or preclinical development. Here, of course, our most advanced asset is EVX-01, which is phase 2 metastatic melanoma. We just released initial phase 2 data towards the end of last year, supporting the very strong phase 1 data that we have seen.
Another example here is the partnership we have with Afrigen Biologics, where we partnered around one of our pipeline assets, EVX-B2, having Afrigen Biologics pursuing an mRNA-based version of that compound. Finally, it's about responders. This is taking our data, our predictive capabilities, deploying that in a slightly different way than what we are doing with the rest of the business, but it's still about saving and improving lives. Here we had a proof of principle for our checkpoint inhibitor prediction model, which released towards the end of last year. Checkpoint inhibitor market is reaching $53 billion in 2024. Think about that: $53 billion for a single class. And checkpoint inhibitor, while gold therapy or golden standard for immunotherapy, is still only a fraction of patients who respond to that.
If we are capable of taking our predictive capabilities, our data insights, and predicting who are the patients who are not going to respond to a checkpoint inhibitor so you can get that patient onto another, more effective therapy quickly, you help the patient, you help society save a lot of cost. You'll also hear more about that later. Look at our strategy: three-pronged business model. It's about targets, pipeline, responders, AI immunology at its core. Pretty simple, lots of opportunities, hard to execute upon, but we are well on our way, as you will also see, later. Another thing that's important and which links into Andreas and Nils being laughed at 15 years ago is the fact that we have been building a multidisciplinary capability set around the platform. On top of that, we have also been building state-of-the-art facilities. Why is this important?
This is important because many of the AI companies today claim that they can do target discovery. They have a couple of retrospective data sets they're working on. We have been building a multidisciplinary capability set allowing us to do constant learning loops, generate our own data, generate proprietary data, quickly test out hypotheses, and continue to develop and feed our model. And if you look at it, I mean, it starts with disease biology, then it spans target discovery, design, the whole preclinical, clinical CMC. It's a strong capability base we've been building around the AI immunology platform. What's also important is just out here, we have our state-of-the-art lab. We have our state-of-the-art animal facility, which again allows the bioinformatics team here to come up with a hypothesis.
We go straight into the lab and test it in the animals, and that gives us a unique advantage compared to the other companies who started doing AI a few years ago. The fact that we have been working on this for a long time also means that we have something which not many others have, which is a clinical validation of our platform. What you see here is data from our EVX-01 phase 1 trial. There are 2 cohorts. The blue one is the patients where our model predicts the neoantigens we're using here, targeting here, are high-quality neoantigens. The red one are patients where the model says these are not so good, quality neoantigens. Of course, any given patient will have a different number of targetable neoantigens of varying quality.
When the model says these are high-quality targets, we also see that that actually leads to a statistically significant longer progression-free survival, meaning that when the model says these are good targets, this is going to work, it also works. So clinical validation of our platform, which is quite unique and also a testament to the capabilities we've been building around the platform. All of this, you can say, sums up to this, that we have a differentiated platform. The validation of the AI immunology platform, the multidisciplinary capabilities, that do put us in a different and clearly differentiated platform or position when you compare it to other companies. And it is something which is enhancing the value of our platform. It is something which is a super strong value proposition towards partners. And it's also something which has materialized in a very strong pipeline.
It's clear that our pipeline, it do demonstrate performance and scalability of our AI immunology platform. I'm not going to go into detail with the pipeline because there will be plenty of opportunity to touch upon that in the subsequent presentations, where we will be going into details, but it is really a testament to the scalability of our AI immunology platform. And then I'm just going to wrap up with looking back a little bit because, how have we actually fared in terms of progressing on our strategy? This goes about six months back, and I think it's fair to say we have seen a very strong execution on our strategy. If we look at the slide here, in September, we announced a collaboration with a leading pharma around discovering and developing a novel vaccine candidate. That's the one where we announced a couple of weeks back. It's MSD.
We have successfully completed the first couple of phases of that and are now progressing into the next phase. Super exciting to be teaming up with a world leader in vaccines in developing a novel vaccine candidate against a bacterial infectious disease with a higher unmet need. We also, as I touched upon, announced a collaboration with Afrigen Biologics. We have been publishing encouraging initial phase 2 data on EVX-01. And also important, we delivered the proof of principle for our responder model, AI responder model for those checkpoint inhibitors. And then what you also hear much more about today is, we've been working around the precision vaccine concept. Super exciting concept because it can, one, help us address patients with cancer where traditional immunotherapy doesn't work. It can also help us broaden the opportunity space for cancer vaccines. You'll hear much more about that during today.
And then the final thing is we've been raising funds a couple of times, first towards the end of last year, and then early this year. In both of these rounds, we had MSD participating, investing, which also means that MSD now is our biggest shareholder with around 15% of the shares. So having a corporate partner working on a novel vaccine, but also same company as a sizable shareholder, of course, also something that means a lot to us and is a testament to what we do. Finally, before getting to Andreas, looking ahead. Plenty of exciting milestones. The last six months have been exciting. The next year is going to be exciting. We will have milestones both on single compounds, on collaborations, on platform, etc.
I will not go into detail with this now because focus is on today and all the discussions that we will be having, but just say that, rest assured, there will be a number of exciting milestones coming over the time ahead. But now it's time to look at AI-Immunology and, who would be better at giving an overview of what it actually is than, I would say, the father of AI-Immunology. Andreas, will you take it from here?
Ja, tak, Christian. So, AI-Immunology. So, AI-Immunology is a platform, it's a true differentiator, and, like I said in the beginning, we have worked with this for many years, 15 years actually.
So, now we really harvest from it, we are ahead of the game, and now I will tell you about, not to go into too much detail about it right now, but just tell what is it and what can it do. So, the AI-Immunology™ platform consists of five different models, AI models, where it has been trained in cancer and infectious diseases. Two of the models, they are within the infectious diseases, that's EDEN™ and RAVEN™, and the other three models, that's within the cancer, that's PIONEER™, OBSERVE™, and AI-DEEP™. So, in the next all the presentations, we will go deeper and deeper in all the models and explain them, but also in the next slide, I will tell you about actually the slide that Christian talked about, that we can actually link, we have a clear link between prediction and patient outcome.
And to me, that is the first time I have not seen any other company showing that, or publications and so forth. So this is truly unique that we have been able to validate our platform in humans. So, like Christian said, this is truly unique that we are able to have a prediction that can link to outcome of the vaccine that we get in patients. But the take-home message here is also that this actually tells us that it's very important that the things you put in the vaccine are the right thing. And what we can do, we can use our prediction score to tell us, is this the right thing? So this is truly unique for us.
Then, just to put things into perspective, before I go too much into the details with the platform, I think you deserve to know the journey, how this started and so forth. So, it started with my mother actually being a researcher within vaccines, tuberculosis vaccines, and that, of course, inspired me. So at that time, in 2008, where the global readiness for AI was really low, I had a vision that I would put the immune system into a computer and then make better vaccines. And like Christian said, that people, they, you know, of course laughed, but they thought it was cool, you know, and I just continued.
So me and Niels, for six years, we struggled and got a lot of non-diluted funding in to prove that the EDEN™ system that I have made actually worked. So it works by faster and cheaper identifying the right components that should be in an infectious disease vaccine. So that worked out, and we succeeded. But then we also went into cancer in 2016. It was actually because of the checkpoint inhibitor invention that you could actually use the immune system to fight the cancer. So that discovery led us to having an ambition of actually making personalized cancer vaccines. So we collected a team of a dream team from DTU, where I come from, with super cool and very, very good bioinformaticians. We built PIONEER™, and after that, we then collected a team of experts within preclinical AI validation.
So at that time, so now we actually had a super team so we could go into the clinic. And we went into the clinic in 2019 with our first personalized cancer vaccine, and where we used the PIONEER™ model to make the vaccine. And that's called EVX-01. That was done at Herlev Hospital. And the next year, we, we moved fast. So the next year, we opened a new study in Australia with our vaccine called EVX-01, EVX-02. We were very blessed because Ingrid Svane, one of the top scientists in the world within cancer, she was the PI of this study at Herlev. And also in Australia, Professor Georgina Long, who's also a star in melanoma cancer, she also led this study. So good, good guys and on board, you know, you can have success.
So we IPO'd in 2021, and then it was really a crazy time because at that time we also had COVID hitting us. So we were actually pushed in the direction of enhancing our AI-Immunology platform to predict virus vaccines. Before that, we have had the focus on bacterial vaccines. So that meant that we created a new model. We built a new model, called RAVEN. Then the two next years, we created AI-DEEP and OBSERVE. So AI-DEEP is predicting checkpoint inhibitor response, and OBSERVE is predicting a new type of cancer antigens that we will present later. And then, of course, the global readiness throughout the years grew and grew and grew. And we are very lucky now that we have partner opportunities like never before, and also have Big Pharma on board.
So we can make vaccines together and have the big guys carrying them out to the patients. So why do we need AI-Immunology? Well, like Christian said, millions are dying from cancer each year, also infectious diseases, almost the same, right? Almost 8 million. So with AI-Immunology, we are able to discover and assess the protectiveness of the antigens. And we are also able to design them within hours, faster, cheaper, and we are speeding up the vaccine discovery process. So that is our solution to this problem. Let me take one step back because, I mean, sometimes we forget what it is a vaccine, and what is the antigens of a vaccine. So a vaccine teaches the body to fight an infectious agent and disease.
It is a drug that introduces to the body, then it can trigger immune response where you can prevent the disease or you can control it. So you can guide the immune system to take the right decisions. A vaccine antigen, that's a part that you put into the vaccine. So that is a part of the infectious agent, where you can train the immune system to learn to attack the bacteria or virus in the right spot. And with AI-Immunology, we are able to identify these components, these antigens very fast and very cheap, and we also identify the ones that are highly protected.
So now I'm going to talk about what is actually the basics of creating a good vaccine with not only the antigens, but also what do you need to understand? So you need to understand the disease. You need to understand the immune response. And you need to understand how to design the vaccine. And all this comes together. You have to put into to connect the dots here in order to make an effective vaccine. And the AI immunology is actually built upon these areas. So we have we decode the disease, we decode the immune response, and we make a vaccine based on our this is actually just code. This is AI coding. So these can be divided in building blocks.
So with AI-Immunology, we have this ensemble with the small building blocks where we utilize them across the models. So we build up like Lego. We build it up. And now I'm going to show you how we actually built one of our, oh, that comes later. Sorry about that. So to go a little bit more in the details. So what we do with our building blocks is actually by the disease decoding. What do we mean by that? We're actually finding the weak spots of the disease with this code. Of course. So we have 13 building blocks here that we can combine when we build the models. Then the immune response decoding, that's where we go in and we understand the immune system, and then we rank these weak spots, which one to attack.
We consider B cell, T cell, and antigen presenting cells, that part of the immune system when we need to rank these components, these antigens, weak spots. There, yeah, we have six there, right? Then we have the vaccine design. We consider safety. We consider to optimize the sequence. We also have models to design the antigens, to remodel it. And there we have these seven building blocks. So now it's going to be funny because now we're going to build a model. So the way that we have built OBSERVE, we first made the first layer. That was to decode the disease. We picked out the important needed building blocks needed in the model. Then we have a second layer where we decode the immune system or decode immune response.
And there, when we want to identify this kind of cancer target, we need these modules, these building blocks. The same goes here. We just build it up. Here we consider safety and the quality of the antigens. And also with OBSERVE here, it's personalized. So we need a personalized building block to put that into the model. Then we have OBSERVE 1.0. So as you can see here, these are the models we have now and all the building blocks in the models. And we have demonstrated that by using the AI models, we have built up a scalable pipeline with pipeline products, vaccines against infectious diseases and cancer. But we can build new models. We can, so what is next?
Well, our building block architecture enables us to scale to other therapeutic areas. And right now we are in cancer, we are in bacterial diseases, and we are in viral diseases. But it doesn't stop here. I mean, what about autoimmune diseases? What about imbalances in the microbiome? What about allergies and what about parasites? So in summary, with Evaxion, we are the first mover of using AI for vaccine target discovery, design, and development. And we have a clear differentiated position. We have trained our AI immunology platform in cancer and infectious diseases, and we have clinically validated it. And our AI immunology platform consists of building blocks that we can utilize across models and build new ones.
And then, like I said, AI knowledge, building blocks modeling system, you can call it, or architecture, it enables us to scale to other therapeutic areas beyond cancer and infectious disease. But just to fly you into the next presentation, one of our building blocks is actually represented across all the models. It's a central building block, and that is, has its own presentation today. And that Michael will talk about that. That's called EVX-MHC here. And Michael has a PhD in bioinformatics from DTU and is the lead architect in RAVEN and in EVX-MHC. Thanks, Michael.
Thank you. Yeah, thank you. So I will talk about EVX-MHC. So it's, yeah, it's really one of our core building blocks. And it's used across all of our models.
The reason that it's used across all of our models is that it's really used across the immune system. So it models a very important interaction between the MHC molecule and the peptides. So what I've shown here is a schematic of a cell and a T cell interacting together. So this happens all over the body, in the immune system, between normal cells, between immune cells, and between T cells, a central player in the immune system. So what is happening here is that the cell is displaying something via this peptide to a T cell receptor. And this T cell receptor functions kind of like an antibody. So it has similar properties to an antibody.
It recognizes this peptide, and it can recognize a good peptide, something that belongs to the body, and it can recognize a bad peptide, so something that belongs to a disease, a pathogen, or a cancer. That is really what is being communicated. The cell could say, "I am diseased with a virus," and show this peptide to the immune system, and the immune system will react, and maybe kill the cell. Or it could be an immune cell that says, "I found this peptide. It is dangerous. We need to, we need to do something about it." So that is really why this is so core in the immune system. The MHC molecule, shown here, is a large protein. It sits on the outside of the cell.
It has this binding groove on the top of the protein where you have this little peptide located. So, what is really crucial here is that this MHC molecule is incredibly diverse. So there are, if we all had the same clone, if we all had an identical immune system, you could just have a virus randomly evolve and then say, "Oh, look, I'm no longer being displayed on this MHC molecule," then it would just destroy us all. So that is obviously not very good. So then we have developed this diverse repertoire of MHC molecules that is different from person to person.
Yeah, so that is, of course, nice from an immune point of view, but it's a little bit annoying from a modeling point of view because that means that we have to model all 26,000 class one alleles, which are the class one is the one that is sort of communicating, "I'm in danger," from a normal cell to the immune system, and 11,000 Class II alleles, which is the one that's sort of the between the immune system communication. So we really need to, yeah, to have a handle on all of these. So how we are looking at this problem. We are taking the peptide, which is a string of amino acids, and then we are decoding those down to a letter. There are 20 amino acids, so we use 20 letters for this.
The similar, the MHC molecule, it's a protein, so it's also composed of amino acids, so we can also decompose that to a string of amino acids. Usually you will compress that a little bit to say that what are the essential amino acids in the binding groups. And then, what is the challenge here is that there is an incredible amount of different peptides that you can have. And there are quite a lot of MHC molecules, so we need to figure out is a given peptide a binder to a given HLA molecule?
The reason that we need to know this is that, if we have a cancer patient that have a certain HLA type, we need to know what new antigens are being presented by this patient, or we're designing a T-cell vaccine, and we need to know which types of T-cell antigens are present in this population. So we need to know exactly what peptides we have to design our vaccine with. To show here is that modeling this actually has some real-world importance. So, Andreas mentioned OBSERVE. I'll mention it again. Christian will talk about it for real later. But we have actually tried this with our previous version and then our new version. This is designing cancer antigens. So what you see here in the middle is a growth curve for cancer.
What you want is the least amount of cancer possible. That is the yellow line at the bottom. And that's our new building block here, compared to our older building block, which gives less cancer, but it doesn't quite cure it. So for certain applications, this really does matter that we keep developing these tools. And the way that we do this is that, first and foremost, we need the data. So there is luckily a lot of data on this out there in the literature and publicly available. And then we take as much as we can, and then we also generate some of our own. And what we try to do is this as cleverly as possible so it matters the most.
And then we look for what are our blind spots in this area. So as I mentioned before, there is sort of a family of these genes that are many different, and they group together. So which ones are kind of similar? So the ones that are similar, if we already have data for these, we don't need to generate more data. And that is what you see on the right. So we have some on the far left of the plot, which are the ones that we have data for. And these, we maybe don't need more data. And then we have some that go to the right. They are far away where maybe we don't have data. And then we also look at another component, which is how often is it represented in the population.
That means that it, it's more likely that somebody will have this allele. So then we generate where our biggest blind spots are. And we try to do this continuously. And then, finally, for this iteration, we've also upgraded the architecture, the modeling approach quite a lot. So we have used a brand new architecture, a deep learning. And we have also made a new training approach called the generative adversarial network, that I'll get back to. So this is how we normally do it. Like I said, we have our MHC molecule on the left encoded as a string of letters. We have our peptide encoded as a string of letters. That is nine amino acids long because that is the amount of amino acids that usually fit into this groove.
And then we feed it into a neural network, which then learns, does this MHC match this peptide? And this works really well. But what we then saw was that we actually didn't get through all of our data before the model finished training. So we actually had a functional model, but we didn't use all of our data. So that puts us in a nice position because we can actually upgrade the neural network to something much deeper and much more complicated because now we are going from a medium data situation to an actual big data situation. So that means that we can make the next generation. And what we do is that we looked at language modeling.
So we do that quite a lot when we do proteins because they are quite similar problems when you put them into the computer. A language is composed of words or letters strung together that makes a sentence that makes sense. And proteins are strung together by amino acids that are put together, and then they sort of make a protein. So very different fields, but the machine learning methods are quite similar. And as you all know from ChatGPT and all the things, there has really been some incredible success in this area. So that is also why we are taking a lot of inspiration. And the main source of all of this success comes from, on the left, this transformer architecture. So that is a way of setting up a neural network to learn these language-specific learnings.
And what it started with was a machine translation. So you have some input, which could be a sentence in Spanish or English. Then you encode that using a transformer. So the encoder transformer, that is then transformed into some a lot of magic numbers. So a large matrix of just numbers, which is then fed into the decoder, which decodes it into a new language. So that could be Spanish or English or whatever you want. So then you learn the structure of the language. Then on the right is just for showing how ChatGPT works. There you just predict the next word. So you write a sentence, what is the likely next word? And that is actually all that's happening in that model. That's quite interesting that it works so well. But all of them are based on the transformer.
So, we chose to use the encoder decoder because it fits our problem quite well. We can encode the MHC sequence. That is sort of the language that we are speaking. So what fits in this groove and what doesn't. And then we use this encoded MHC sequence, together with the peptide. And then that is decoded into a prediction. This is a binder. This is a non-binder. And this should work really well. But initially it really didn't. It just never, ever gave a right answer. It just gave complete nonsense. So that means that we had to take it to the next level to actually train this model to make sense. And the way that we did this is through something called the generative adversarial network, as I mentioned before.
So this means that we basically take the network and then we duplicate it. So now we have two networks. So one network, it makes fake peptides. So it basically deepfakes a peptide. And then we have another network that it learns whether or not this, the input is a deepfake peptide or a real peptide that we have from our data set. And that means that these models can learn together that it's a more stable way of training them. So, so initially you just randomly initialize your models. So that means that the output will be complete garbage. So that means it's really easy to recognize, is it a fake peptide? Is it a real peptide? That's very easy.
So you learn slowly, but then as you learn to recognize the difference, you all the generator, the ones that make the fake peptides will also learn to make better fake peptides. So therefore you train together these models. But that also means that once you are finished with this pretraining step, you have a model that really knows what are the intrinsics of a peptide. So they had a sort of a deep understanding of what is a peptide. So that means that we take it to the actual training and train what is a binder, what is a non-binder. It has a much more stable and improved way of doing it. Yeah, so we are training together class one and class two.
Normally you train these separately, and then we fine-tune models using only class one data and only class two data, in the end. So what we get is these curves. And they show gradual progression. So you want good recall. That is how often you find your positive. Then you want good precision. That is how often is what you say is positive, actually positive. So we want to go to the right and we want to go up. So that is what the perfect predictor would be a square in this plot. So, as you can see, we have our beginning, our starting point framework. And then we add this transformer and then we get better. And we get better for both.
and then we add this generative adversarial network and then we get better for class one and we get a lot better for class two. So you can really see that the area between the curves, it's quite large. And that is really significant because it's a much more complicated challenge to predict these class two bindings. Yeah. And we can also prepare. So like I told you before, there are thousands of different alleles. So each dot here is representing one allele. So as you can see, there is quite a difference between the performance of the different alleles. So that depends how much data and so on.
But we have really pushed a lot of the alleles up and we have really the ones that are at the very far bottom, we have also taken those quite a lot up. So compared to ourselves and also compared to what we have, what you can find in the literature and what is freely available on the internet. Yeah. So, like I said before, I'll reiterate, this actually has real-world implications. We can in our OBSERVE model see that the antigens that we design using our newest model, it's much more responsive in the immune system, react to it much higher. If we look at the neoantigen immunity, we have a higher precision. So we have more hits when we use our newest model.
So that is really why this is used across all of our different models. And finally, of the peptide MHC, it's really key. I would like to, of course, reiterate this. And it's really crucial for modern vaccines to get this right. So, for some vaccines, T cell vaccines, the cancer vaccines, you need to get this right. There are no other options. For B cell vaccines, it really helps. And our newest iteration, version four of this building block, it's quite a significant improvement in this field. And that actually improves real-world vaccine performance. So thank you.
Thank you, Michael. And I think this is quite unique, right? And that also speaks to this, you can say the unique architecture we have of our platform where we have these building blocks that we are using across the platform. So that means when we come up with something absolutely brilliant, which increases predictive powers, we just plug it into where it fits in all the different models. And then we have a major benefit here. So I think that's quite unique that we have this building block approach here.
So with that, we have time for Q&A. And those who are online, please just type in questions in the chat. And those who are in this room, please just raise your hand in case you have a question. Klaus.
Would you try to explain the level of, let's say, precision that is needed for enhancing the efficacy of a vaccine?
Yeah. So that's a very good question. So it also depends what you have available. So if we're looking at something like a pathogen, then you have thousands of peptides. You have a lot to choose from. So then it doesn't matter too much how much your precision. So there you need maybe 75% precision or something. So but if you have something like a cancer or an ERV, there you only have 13. So you really need to get those right. So going from what we normally see, 80%-90% or 95%, it makes a hell of a lot of difference because one more hit that makes a difference of whether it's an efficacious vaccine or not. So we need to chase those improvements in that area.
And Rasmus, I don't see any questions here. Do you see any?
I don't see any.
That was because it was perfectly clear. Yeah. No, we know it's a little bit technical start, but it's also so central for the whole platform that it's worth spending time on. And because that is really what is helping us improving and or saving and improving lives, that you have a building block which can increase the predictive capabilities beyond what you can do with any other tool out there. So it's super important. Question online.
Oh. Richard, any comments on Moderna's vaccine and how does your technology compare to theirs? Maybe we will cover that in the PIONEER section. Or do you want to?
I don't know which one that is. I can look it up. Okay. But we'll take that in the PIONEER section. Klaus.
Could you try to explain what kind of data set would you need to enhance your model's efficacy and prediction? Would that be external real-world data or animal models or data?
Yeah. So how do you normally generate this model is you look at a cell and you see what it presents. So then you can get a lot of data out of, but you don't get the disease state of the cell. So more data that models what happens when the cell is, like, in danger. That is the type of data that really improves these models.
Well, I think one important point is also because when I joined, I was kind of like, why don't we talk about how many data we have, right? Then, I kept being told by the team, well, that's fine, but it's more the quality of the data and the different data sets, right? So I think it's getting that right composition. That's, of course, also the fact that we can generate a lot of data ourselves here gives something quite unique. So it's not about how many terabyte, but to some extent, but it's also the composition of data.
So is it more the description of the cell and the knowledge about how cells actually function that can enhance the vaccine design and the precision?
So are you talking this specifically or vaccine design in general?
In general.
Yeah. So then it's it is sort of picking it apart, right? So there are some blind spots where, yeah, we don't know really how what happens when you have these disease states or other parts of the interaction. So those are really incredibly difficult to model. So it's also a question of where do you use your resources? Because there is one spot where you could have an incredible improvement if you could predict this, but also you would need billions of data points for an assay that don't exist. So that is just not really feasible. So you need to really spend your resources where they actually matter and where your model can fit them.
Yeah. Thanks. I have specific questions for both speakers. So I was just wondering about the phase one data. It was interesting to know n=6 means n=6 patients per group or 6 monochromes per group?
6 patients.
And then you gave a highly predictive antigen with a high score and then an antigen with a low score.
Yeah. And no placebo. So that was the best, yeah, because that was the best vaccine that we could create. So we couldn't find any more optimal neoantigens based on the algorithms we had. So and after that, we saw that there actually was difference on the protection. And this also comes back to the question also from Richard, also Moderna and so forth, that it's really important the quality of the neoantigens that you use. And I know, I think, but that's my remote memory. I think it's 20 neoantigens are used in the vaccine. Rasmussen?
34.
34? Okay. 30. Yeah. It's a lot, right? But here we have another approach. We only have 10 in general. But we show that it's extremely important that you have the right neoantigens and that we can show with our prediction scores because we can link the outcome of the clinical trial with a prediction score. Thanks.
And then we have, Lee, a good question. How do MHC predictions translate into CD4 and CD8 reactivity?
Yeah. So that is also a big area of research where we are going because, of course, you, the very first slide, there is a third component, the T-cell receptors that also need to react to these peptides. And modeling this is incredibly complex. There are billions of TCRs. You need to say that the peptide MHC binding is a prerequisite, but it doesn't mean that you have reactivity. We are doing the best we can to model this, and we can improve a little bit, but it's still sort of an open area. So you need this binding, but that doesn't guarantee reactivity.
And then we have another question around the announcement from yesterday that the Novo Nordisk Foundation and EIFO and NVIDIA is going together and creating what's one of the five biggest supercomputers in the world, and what that means for us. I think I wish we had timed this event for this event here. But even though many of us know foundation quite well, that's not what we did.
But it's, of course, obviously very good news, because it is going to be something which is going to be supporting the life science world in Denmark and and AI. But I think what it practically means, probably too soon to say, but definitely an important step forward in advancing and creating an even more supported life science environment in Denmark when it comes to to AI. And then, Michael, what is your gut instinct about the progression of EVX-MHC? Where are we at, 4.0? What are the quantum leaps that we can can expect?
I mean. If we can get any better. There's always room for improvement, and especially this class two. So, so for instance, the whole field has been kind of neglecting, so we're using the one called DR. So but there's also DP and DQ, and they have been a little bit neglected.
But I think with the performance we're seeing, we can now actually take those into account. And the reason why they are, there's two of them coupled together, so it makes it exponentially more complicated. But we can slowly start to see the performance where we can actually use these. And of course, we're going to keep training. We're going to keep using new models. And I can also see there's some questions about diffusion models. We're definitely also going to look into those and see if that can improve our performance. So the whole field of deep AI is evolving extremely rapidly. And that means, of course, that we can just take into all these architectures and techniques and then just try them out and see what works.
So once we have something that works, then we, of course, dive deep and then make it really work well, for our problems. It's really exciting.
Then we have, final one here. Have you tested the possibility of using your AI models to design and test also peptide ligands that can include or augment responses to cognate peptides that are naturally recognized? Something we,
Not yet, no. But that is definitely also very exciting.
Excellent. I, I think, quite clear that, super important, building block, which is crucial for, the whole platform and, quite, unique predictive capabilities. But, and thanks for, for the questions here, because in the interest of, all of you getting a break as well, we'll, end the questions here, and then, we start in, a little less than 15 minutes. So, thank you, to you, Michael, for taking the questions. Okay, take a seat.
We're back from the break, and I don't think I need to introduce the next speaker because you have already seen him, the father of it all. So I'll hand over without much further ado. Will you take it from here? Yes, I would love to.
Now, thanks for a nice break, and nice to see you again. So I will present, yeah, in the next session here, EDEN, my little baby. So EDEN is best-in-class, best-in-class model, in assessing protective B-cell antigens. B-cell antigens are those components that create antibodies in the body. So why do we need EDEN?
Well, the global deaths from antimicrobial resistant infections is skyrocketing. So just to give you some numbers here, today, 10 million die of cancer each year. When you look at the antimicrobial resistant bacterias, it's 1.3.
So it's a subpart of what we explained in the last session, where it was 8 million in total in infectious diseases. This part is 1.3 today. But it's predicted that in 30 years, we will have the same amount of deaths as cancer in antimicrobial resistant bacterial infections, as we have with cancer today. So there's a huge need of doing something. We can't let this happen. So we need AI for accelerating the vaccine development against these bacterial infections. And vaccines have a proven track record of combating the infections. With traditional vaccine development techniques like reverse vaccinology, you can find the right antigens, but you have to be lucky. It's very expensive, and it takes a long time. And I will go into details with that in some slides where you can see some examples.
We use EDEN™ to rapidly identify the antigens, but we also find the ones that are highly protective, and they are very broadly protective. I also showed that later, so we have a solution to this problem. As I showed you in the last presentation, we have built EDEN™ with these components. So as you can see, some of the building blocks are B-cell antigens. Some of the building blocks are Evaxion Biotech , as Michael talked about. And then you can see that we actually are targeting both virus and bacteria. So how does EDEN™ work? EDEN™ identifies the novel protective antigens by feature recognition. So what is that? It's actually like face recognition. You use the features of a protein, you train on these features, and then you learn how to find new highly protective proteins.
So EDEN™ has been trained on highly protective proteins, very unique highly protective proteins, to find new, very unique, highly protective proteins that are novel. One feature could be the surface of the pro, of the, of the antigen, that are likely to be, where an antibody can bind. That could be a feature. That's actually one of the features of EDEN™. But there are many features in EDEN™. So how does EDEN™ work? It works by taking in the proteins, all the proteins, all the amino acid sequences of a bacteria's proteome. A proteome is actually just all the proteins in a bacteria. That's called the proteome. So all of these protein sequences are fed into EDEN™. And here you see, this is just an example. There's 2,607 proteins fed in. Then it calculates for about 24 hours.
Then the output is a list of each protein, what is the predictive score on this protein should be protective or not. So as you can see here, the score is from 0 to 1. So in the top, the best protein here that EDEN has found is ranked as number one, and that has almost one as an EDEN score. So what we do here is in Evaxion, we then process the proteome of a given bacterium we want to make a vaccine against. Then we get the ranked list. Then we look in the top, and then we take these potential highly protective proteins. We express those, so we have them physically, and then we test it in the lab for protection. So we have an AI score, and then we have a reality check if this is, that's the way our Evaxion protection.
And that I will show you later, actually, is really remarkable. And first time ever it's done in the world, actually, shown in the world, to my knowledge, that you can actually rank protectiveness in real life with the actual AI score. And that's what you see here. So the EDEN AI score correlates with protection, and we can use it to identify not only rapid and with low cost, antigens that are an ask, but we can actually see, okay, we're actually getting proteins that are highly protected. So this is the top rank. This is the top rank of each of these four bacteria that we have tested. So not the bad ones, but the ones that we found to be highly ranked. And here you see a significant correlation on all the different bacteria.
So it's gonorrhea, Acinetobacter baumannii, Klebsiella pneumoniae, and Staphylococcus aureus. So the two in the middle are maybe people don't know, but at least we know Staphylococcus and gonorrhea. So this is truly unique. We actually have an AI score, and then we have reality, and we see a correlation. So this is finding the, before it was finding the best antigens. This is, we can outperform reverse vaccinology doing things faster and better. We can nail down and filter down, instead of testing tons of proteins like in reverse vaccinology, we only test few proteins. And this is a study done by Novartis. It's a reverse vaccinology study. So what they're doing is that they test hundreds of proteins.
It took them many, many years to make this, and they ended up having a pool of proteins, those ones here, and tested them in animal models. Before that, they did a lot of analyses and so forth, but then they found these proteins and tested them out. And the red ones are the ones that are highly protected. Then Intercell did the same. This is Streptococcus A, bacteria. They did the same. They also found some of the same protein as Novartis. It also took years of work, and they found additional one, as you can see, one red protective protein. Each dot is a protein. So what if we just run Eden? Well, in the top rank of Eden, all the proteins that have been predicted, we found in the top 40 rank, we found all the proteins they used years of work to identify.
We found that just by running our EDEN model on it. But what we also found was 23 completely novel potential protective proteins, never been tested before. So just to recap, we took the proteome of the bacteria of Staphylococcus aureus, and we ran it through EDEN. It took us 48 hours only instead of years, and then we pinpointed the ones that were protected. So eight proteins was found in top 40, but we still have the additional blue ones that we could test out and see if they are protected. Okay, so how do we use this?
We have used EDEN to solve a problem, the gonorrhea problem because at least we believe that it will be a vaccine that works, of course, in humans, but at least we have found the right antigens. But first, I would show you here the numbers because gonorrhea is skyrocketing, is becoming antimicrobial resistant, and people die of it. So gonorrhea is caused by a bacteria called Neisseria gonorrhoeae. And it's the problem with gonorrhea is that it actually is not often that you have symptoms. So these numbers are only the numbers that are with symptoms, right? So you have a lot of the populations that have actually gonorrhea, and it's a huge problem.
So that's also why CDC have treated the gonorrhea as an urgent threat and that we have to do something about it, because people, as you can see, are dying from sepsis, and babies are born blind. You got an infertility, you got the damage to the nervous system and so forth. It's doing a lot of problems, and that is rising now. So what we did, we used EDEN and said, "Okay, we want to make a gonorrhea vaccine." We processed the gonorrhea proteome through EDEN. We then took the top 30 of the highly protective, predicted proteins, but only the predicted proteins. These are just predictions. And then we expressed those, and then we tested it out at UMass Chan Medical School, with our very good collaborator, Sanjay Ram. And we tested out each protein in the top 30.
Then we saw that reality actually correlated with the prediction score. So it was a huge happiness when we saw this because we now knew that these two proteins that we have in our vaccine, EDEN-1 and EDEN-2, the two antigens, they are highly protected. And we then have made a vaccine where we put the two antigens together. It's called a fusion protein. And then we have one drug. So we halve the price of the production. So but also, as you can see, that EVX-B2 has a higher, more robust protection than the other two. So EVX-B2 is comprised of EDEN-1 and EDEN-2. So two proteins, two antigens together.
And another thing was, okay, now we have this candidate, now we have this vaccine, now we want to test the broadness of it. I mean, because what EDEN™ did was that it analyzed the broadness of the, of the vaccine also. So we actually have a vaccine candidate with these two antigens in the vaccine candidate that are able to kill. And this is the biggest, to my knowledge, biggest, clinical strain, bactericidal killing study ever done. So here you see that actually the vaccine can kill the bacteria, all the bacteria, the 50 bacteria. Normally when we see this, bactericidal killing assays, you only see maybe half of the killing. So, so the question is, let me move your way. So the question is, how effective is this compared to other candidates out there? We have to benchmark it.
So what we did was that we knew that GSK and LimmaTech Biologics and Griffith University together are working on a gonorrhea vaccine. And they are out publishing a lot and going on conferences, and they always talk about this protein, NHBA, a protein that is also an antigen, so it's a vaccine candidate that is known to be highly protective against meningitis B, which is also a Neisseria bacteria. That is a component of the Bexsero vaccine. So in meningitis, this protein had high success in the Bexsero vaccine. And GSK is also working on this protein. So we, of course, wanted to benchmark our EDEN-1 and EDEN-2 towards this highly protective protein.
And what we can see here is that our EDEN-1 protein is more protective than this super antigen. And our EDEN-2 is as protective as it. So we believe we have a really truly unique and very powerful EVX-B2 gonorrhea vaccine. This EVX-B2 vaccine, that is the one that we have made our partnership with Afrigen Biologics that are now testing it out in mRNA. But okay, why is this so protective? I mean, because actually we, when we looked at the prediction, we saw that these proteins in the top rank, we saw that these proteins are actually cell division proteins. And Sanjay and me said, "Okay, wow, wow, okay, why does EDEN tell us that this is a protective protein? It shouldn't be.
It should not be on the surface because of cell division." But we said, "Okay, let's try because the AI system is telling us this is a protective one." And then we tested it out, and that is the most protective protein that we have ever seen. So why is it so protective? It's actually because it's not inside the cell all the time. It actually goes out of the bacteria on the surface. And that is why, but that is why we can attack it. And when it goes out is when it's dividing. So the bacteria is dividing and it's vulnerable in that, so this is a bacteria that is dividing in two. And then we hit it there and then we kill it. So it's a bit like Star Wars. If you know the empty spot, you go for it.
The empty spot, the weak spot, sorry. If you know the weak spot, you go for it, right? So, this is actually an AI-made picture, that we made because then we couldn't because AI solution presented. So, the Death Star looks kind of over. So, but this opens up the potential of, you know, this is a new class of antigens, and we see, we see this type of protein also in other bacterias where there is no vaccine available, whereas there's a huge medical need. So we have, now we have the really options of going broader with this and make partnerships. And, yeah, this is just that EDEN™ is behind several of our infectious disease product candidates, as you can see. So we really utilize the effectiveness of it.
In summary, EDEN™ correlates with protection in identifying the most protective antigens and, potentially, potentially enabling lower risk because you have a highly protective protein in, in the clinic. EDEN™ outperforms reverse vaccinology for faster and cheaper vaccine discovery. We have a solution to gonorrhea with our EVX-B2 vaccine. We have our unprecedented, unprecedented bacterial killing with our, our vaccine, our B2 vaccine. We have benchmarked it against the leading antigen and it outperforms it. EDEN™ has helped us to find a completely new, new type of, of antigens that we can use broadly. And you, Andreas, and while you think of questions, oh, you don't need to think.
You have a question already? I do. Yeah, let's get those two questions.
Yeah, so first, do you have any features that account for fold structure of the protein, solubility, how easy it is to actually express it and purify it, and producibility of the protein? This is the first question. Second question, when you say percent protection, is that based on cell-based assays or animal studies or?
Yeah, that's a really good question. So the first one, so this has been tested by UMass Chan Medical School, Sanjay Ram, in a highly advanced sexually transmitted mouse model. So it's a really unique model. Only one lab can do it in the world, actually. It's percent protection. Yes, that's percent protection when you have reality of the, of the pro, of the antigens, how protected they are. So that's the sexually transmitted model.
Oh, yeah, okay.
So you actually, what you're doing very, very quickly, you're actually giving them antibiotics. So you remove their microbiota, and then you go in, and then you push them with hormones so they got estrous cycling through estrous cycle, yeah, exactly. And then you can actually infect it with the gonorrhea bacteria. And that's what we show. So it's actually a clearance model where you see the how fast it gets cleared. So that's a protection.
Then we have a question here. How do you make your prediction based on the proteome with EDEN? What are the predictions based on? What do you match them with?
So we take, in gonorrhea, we took all the proteomes that we could find, and we fed it to EDEN. And then we calculated something called a cross-protection score so we get the broadness of the vaccine. So it's about also having this conservation in. And then, when you mean protective, what kind of experimental models were considered to predict the protectiveness against of the protein against the bacteria? Yeah. So, gonorrhea, that was the sexually transmitted clearance model. But the other three bacteria, that was survival. So it's actually just a sepsis model where you vaccinate it, and then you see if it survives or not. So it's a more bound model.
And then, a little bit along the same line, how did you train the model? And this is in the reverse vaccinology example. Assume you did use the data from Novartis Intercell publications, or you did not use the publications, so it was just based on the already trained EDEN model, right?
Yeah. So some of the proteins that EDEN have been trained on, why coincidence was that a couple of these proteins that were found were also trained in EDEN. But when we then did the prediction, we left it out. So we have, so you should imagine EDEN as a system where you have proteins that you train on, you have a lot of proteins you train on. But when you then want to predict on another bacteria, and if you have used, like, Staph aureus proteins, and you have used that to train, and you want to predict on Staph aureus protein, you don't check that train in EDEN. You don't use that component on EDEN. You actually, so EDEN has a lot of trains where they, so it's not biased to what it's predicting. Because else it's just predicting what was trained on.
So it's important to leave out the proteins that you want to predict on, or the whole bacteria type. So that's how we do it.
And then, looking at the competitive landscape, are there similar AI companies developing AI-predicted protective antigens, and can they also do it within 24 hours?
I have not seen other companies presenting this. I know that Oxford University are having some academic prediction tools, but they have not talked about how fast they can do it and so forth. So not to my knowledge.
And then I see there are quite a few questions. We'll take a few more, and then we'll ensure that those we didn't get to answer now, we will follow up with answers on those. And I don't know any questions in here before we take a few more here?
One more question. I mean, you didn't answer my first question.
No, the first one, yeah. That was, sorry about that. That's an old trick, you know.
I noticed.
Sorry about that.
But producibility, so you can, you know, okay, there's protein sequence. That's what you, that's your output, right? Protein sequence. How do you know if that you can produce or not?
Yeah, yeah.
If it's still cellular.
Yeah, exactly. And that's a very good question. So, so when, when we get the output, you have, a protein as a full-length protein. But sometimes this protein cannot be produced, as you're saying, because it's a, it's a, can be sit in the membrane and it can have some elements that are not reproducible.
So what we do, that we use additional bioinformatics analysis tools, and then we go in, and that was also in our building block vaccine layer. We go in and then we remove these elements that can't be produced, and then we put it together, we knit it together, and then we have a structural prediction building block that go in and ensure that it's still the right for folding and the right one. And there AlphaFold is really amazing.
Thanks for that. And then with regards B2, do the two antigens target the same or complementary biological pathway? And do you see synergistic potential of combining the two?
Yeah, that's a really good question. So the ones that were the most protected, you could say, better than the other GSK protein, that is a true cell division protein. The other one is also a cell division protein, but that is called [uncertain]. And that also has some; it has dual functions, so it's also responsible for invasion. So it's actually used; this, the bacteria use this protein to come into the cell, the human cells. So they have the share of the functions, but they're also different. So synergistic effect, yes. I strongly believe that that will be the case. And then, one other question, which is, how come EVX-B1 and EVX-B2 don't have a corporate partner? And you can say,
of course, for EVX-B2, we do have a partner, Afrigen Biologics, for low and middle-income countries, but we don't for outside that, and we don't for B1. And why don't we have that? We haven't been that active in partnering, and that's of course a key focus for us now to advance our assets via our new multi-partner approach. So, we are out there talking to different parties. And, once again, thank you for all the questions. I'm sorry that we can't get to answer them all, but we need to get onto something just as exciting as what we have been. And will you introduce the next topic?
Yes. So as you can see, this is Raven, and Raven is going to be presented by Michael again. So, the key architect here from Raven. So thanks, Michael. Thank you. Yeah, so I'll talk about Raven.
So, Raven is our model for making T-cell antigens, something that we think have been quite underappreciated in infectious disease vaccines.
So really, our key goal here is to distill down the immunogenic component in the whole pathogen into something that is easily produced and very immunogenic. So for instance, the big spaghetti here is the spike from SARS-CoV-2 virus, the common vaccine component that is in all the different vaccines that are around. And it works really well. And the reason that it works really well is that it has really a lot of T-cell antigens. So that is what is highlighted here is just one T-cell antigen for one person. So, like I said before, we all have different HLA, so we're all going to respond to this quite differently. But there's not always a lot of antigens in one of these important subunits.
And there's also the whole concept of you also need T-cells for different parts of a vaccine. So we want to include the whole pathogen. So here we show the whole genome of the SARS-CoV-2 virus. The purple is the spike. It's a small component of the whole virus. So what we also would like to do is to take out the crucial components. We call them hotspots that we can then show the T-cells and have a really efficacious vaccine to boost even more. And so what we call a hotspot is really just a collection of T-cell antigens. So they're going to be, again, different from person to person. So we really need to make sure that we have a good collection and we have the right connection.
So that is what this will do. So, as I've also told you, that you really need T-cells to get the B-cells going, which is the antibody-producing cells. So if we provide the right T-cells, we can really boost the antigen response. Yeah, so that we have also shown here. And it's just a quite nice small study, where we have two different primers. We have the classical spike protein, and then we have two T-cell vaccines. So they are only some of these small components that we take out. And we made two of them, just to show that it works twice. And then we boost with the spike molecule.
So the interesting here is on the left, on the right, is that the antibody titers they go up to a similar level to the normal vaccine. And they catch up using only a single shot of the spike protein. And you might ask, why then do something that is already as good as the other vaccine? But the really key feature here of these T-cell vaccines is that they're really, really easy to make. So we can make them in DNA, we can make them in RNA, we can make them in DNA. We make them for our personalized trials. That means that we make them really fast, and we make them, like, we need them. Okay, we make them in a couple of weeks. So they are incredibly easy to make, and they work across different modalities.
So that is why there are a lot of potential in having these T-cell vaccines. One of the potential here is in a corona-like situation, where it takes a while to get the spike virus in production, and it takes a while to actually confirm that it works. But we know that the T-cell part is going to work. So we start making that, and then we start production on the SARS-CoV-2 virus. So once we have the vaccine, the B-cell vaccine, we already have primed the immune system for this vaccine. That means that the response this time is going to be drastically faster. Then we have also seen that we know from cancer and viruses that we need to kill a T-cell.
So that's a separate part of the immune system, and they surveil cells that are infected or cancerous, and then they kill them. So we also, what if we only target these? What if we don't have any B-cell component? And first and foremost, we try to find these hotspots. And then here we targeted a mouse because that's what we're going to test it on. And we saw that we got quite a good hit rate, what we predicted, we got almost all of them, a response against almost all of them, which is really nice. But also crucially on the right, we actually see that these mice that are vaccinated with this, they can survive a lethal challenge. So they do get sick. It's not as a greater vaccine as a B-cell vaccine, but they are protective.
And this is a vaccine that we can make in weeks. So, as soon as any virus comes, as soon as we know the DNA, this is really fast to do. Then we can start producing a vaccine. And so far we have done this many times. It has worked every time. So you can always produce the vaccine. That's really the key to these vaccines: that they are so fast. Yeah, so we have a model, and then we have different ways to utilize it. So we have the standalone T-cell vaccine that I just talked about. It is protective. It is not the best vaccine, but it is very fast and it works very well.
Then we have the prime combo with a B-cell vaccine, which you can sort of prepare. The response time is a little slower, but it's still quite fast. And then finally, we are also developing into the model that we can now graft these T-cell epitopes that we find into the B-cell. So we have like a one-structure vaccine, a more classical vaccine that takes longer, but it's really high in protectiveness. So the way that this works is that you pick your viral genomes. Sometimes it's just one virus that comes up. Sometimes it's a whole family of viruses. So you bunch those together, and you run them through RAVEN™. You find out what type of epitopes are there. Then you decide on a target population. This we look up in a database.
So we know all the many thousands of HLA types. And then you look what, how they are distributed in this population. And then you have an extremely large matrix of millions of potential antigens. And then you, and then the key thing here is really selecting the right hotspots that are complementary to each other, whereas they're not going to respond to the same antigens as somebody else. So you need maybe some for me, some for, for somebody else. And then if somebody else is also responding to mine, so you sort of have this extreme combinatorial space that you need to distill the virus down into the right hotspots. So, so that is the, the, the key change. Yeah, so, so we have looked at this in the hindsight of 2020. So we have looked at what people have discovered in SARS-CoV-2.
So that is the peaks going up. And then we have looked at what predictions did we make on day one of the pandemic. And we think they line up quite nicely. It took at least eight months to find the purple peaks. And it took us a day to find the blue peaks, that is going down. So it really is a fast response platform. That's how we initially designed it to be. And then we have, now that we're on the other side, we have started looking into this structural modeling. How can we take these T-cell epitope vaccines to the next level? And then you choose your B-cell antigen. And we're just seeing how great EDEN are at choosing B-cell antigens. So that's a good starting point.
All you maybe sometimes just have one. So in terms of SARS-CoV-2, you really only have spike that is going to be good protection. So that, that's your, that's what you have to work with. So that is also what we designed this from, from viruses, where you should just have the, it's a very limited selection, but we are also seeing indications that it's working quite well in the big space. Yeah, and then we simply insert these T-cell epitopes from either across the genome or across the whole pan-genome of all the different viruses into one antigen. And how we're doing this is through a structural modeling tool. So this is also something that we developed a while ago and are still actively developing.
So it's an autoencoder, generative model that looks at small structure components and it encodes them into this. It's called a latent space. So it's, again, just a big matrix of numbers, but it captures some very crucial things that you cannot see when you look at the structure on a computer or using your human brain. It's something that only the machines can see. And that means that we have a bigger space to graft these peptide cells. So we can really make some extremely efficient vaccines using this technology. Yeah, so this is used in some of our newer products. And it's, yeah, so we're really exploring how we can use these. So it's only a few components that you need to add to really improve the efficacy of the vaccine.
Yeah, so we are finding these hotspots, and we are, we're really utilizing our capabilities in the personalized space to make fast vaccines. And we can also see that we can choose to target specifically T-cell epitopes that really boost B-cell vaccine candidates, and make a really great antibody response. Yeah. Thank you.
Thank you.
Then, Michael, thank you so much. And I think it's also clear, not just having EDEN, RAVEN, but the combination of the two offers quite unique opportunities for combining and really making a difference in infectious diseases. So, and it's really an important important asset. And then, here, does RAVEN take phase variation of antigens into consideration?
Variation. Yeah, I know. Yeah, so, so we're not, oh, that's EDEN's question. Yeah. So it's RAVEN. So I think it's just the question of basically the base phase.
Yeah, I don't unfortunately, I don't know what phase variation you're supposed to, but maybe we can talk later. Yeah.
Can RAVEN predict antigenic drift into account for viral antigens?
So yes and no. So we're really looking at the families of viruses. So we can see what has been, and we can identify places where you have a high variability. So then we can target the conserved sites. So of course, when you put an enormous evolutionary pressure on a pathogen, it's probably going to evolve in a way that you might not be able to predict. But at least we can hedge our bets as much as possible, using like any technology that we can, you know.
Any questions in the room here? Good. Thank you, Michael, for both presentations.
What we're going to do next is we are going to transition into another part of the business talking about the personalized cancer vaccines. But before we do so, we are going to take a break and be back here in about 15 minutes. So, enjoy the break, and thank you again, Mike. Thank you.
So thank you so much for coming back here, and thank you so much for actually participating. So we have really a lot of big audiences here today, so investors, family, friends, and partners, and so forth. So thanks a lot. And also on the web, we have a lot of participants, so thanks for that also. It's actually our record, so that's nice. The next speaker. Now we're going into personalized cancer vaccines. And the next speaker is Thomas Trolle. He's a PhD in bioinformatics, one of the dream team founders, you could say. So Thomas is the key architect of PIONEER, and Thomas is going to present this now. So thanks a lot, Thomas, for jumping in. Thank you.
Very nice introduction to the dream team. Yeah. So I'll be telling you guys a bit about our PIONEER model, which is our model for designing personalized cancer vaccines based on neoantigens. So in the past 10-20 years, there's really been a revolution within the treatment of cancers, and that has been based on these cancer immunotherapies using checkpoint inhibitors. These checkpoint inhibitor treatments allow physicians to treat certain cancers that were previously basically untreatable with, for example, chemotherapy. And a good example of this is melanoma, where there's been a huge advance in the quality of care we can give to patients.
However, when we sort of dig into the clinical trials that tested these checkpoint inhibitor treatments, we actually see that it was only around 20%-30% of the patients in the trials that actually got a benefit from the treatment, so meaning that their tumors shrank during the treatment. This highlights the need for improved treatments within this area. Neoantigen-based cancer vaccines are sort of the next logical step in the evolution of these cancer immunotherapies. They are this because they arise from cancer-specific DNA mutations, and this means that neoantigens are found specifically within the tumors of patients and absent from their other cells, from their other tissues.
The fact that they are found specifically in their tumors allows a neoantigen cancer vaccine to elicit strong and highly specific immune responses directly towards the tumor, and hopefully avoiding a lot of side effects that we see from current cancer immunotherapies. We also believe that the fact that we are triggering these highly specific immune responses allows us to synergize well with the current cancer immunotherapies, which is a big advantage when you're looking to validate these in the clinic. There is, however, a challenge in developing these neoantigen cancer vaccines, and that is the fact that each and every patient has a distinct set of neoantigens within their cancer cells. Here I'm showing you a graph from a study where the authors analyzed the genomes of more than 300 tumors from melanoma cancer patients.
What you can appreciate is that the number of mutations, and thereby potential neoantigens, varied a lot from patient to patient, with some patients having less than 100 mutations and some patients having several thousands. We can already see here that, I mean, the set of neoantigens in each patient must differ. They also dug deeper and looked to sort of see where there are some mutations that maybe were found in several patients, and there were a few, but they were very few and far between. There was not a single mutation that was found in each of the melanoma patients. Bearing in mind that this is only melanoma, and if you move to lung cancer and colon cancer, the mutational landscape could also look very different between different cancers.
So this is a bit of a challenge, you could say, in developing a universal neoantigen vaccine. It's clear that sort of a traditional vaccine development approach, where you have one vaccine that is relevant for all patients, that's probably not going to be the most feasible way to continue in this space. So what we think the next step is to tailor the treatment more specifically to each patient. We're already seeing these precision medicine approaches being used today to treat cancer patients, where the doctors are able to measure various biomarkers on the tumors of patients, and thereby direct patients to one or more different treatments.
So with the PIONEER model, we're taking this to the very extreme, you could say, and we're basically designing for each patient a specific cancer vaccine that will fit his tumor specifically, his or her tumor specifically, and their immune system, of course, specifically. As Michael very nicely introduced, there also can be differences in how our immune systems react to vaccines. So Andreas showed you a bit earlier about how we built PIONEER using these AI immunology building blocks. And I'd like to just dig a bit deeper and show you how we actually came to deciding that these were the building blocks that were needed for the PIONEER model. So we designed PIONEER to model mechanisms that go on inside cancer cells that would lead to the generation of effective neoantigens for a cancer vaccine.
So the first step is, of course, to identify the mutation within the DNA of the cancer. That is sort of the whole basis for this working. Next, we look to see which of these mutations are within a gene and expressed in the tumor, which of these mutations then actually generate a neoantigen, which of the neoantigens are then presented on MHC. Again, as Michael went through earlier, this is a critical step in order for a neoantigen to be effective and to be able to interact with the immune system. And here we, of course, use our EVX-MHC building block that Michael very nicely went through in the previous session. Next, we have some methods for predicting which of the presented neoantigens are likely to elicit a T-cell response.
And finally, up until now, the first five steps, we've sort of been looking at a single cell, but now we zoom out, looking at the tumor as a whole. And it's a known fact that tumors can be highly diverse in which mutations they have in their genomes. This is due to the way they evolve inside the body of the patients. And here we also have a building block that allows us to look at the clonality of the neoantigens and select those that are present in all tumor cells, basically. And finally, doing all this, we get our ranked list of neoantigens, and then we use our vaccine design building blocks to sort of design the personalized vaccine for each patient.
So right now, I mean, PIONEER™ is, of course, just one part of a much larger process that goes into creating these personalized cancer vaccines for patients. So here, for example, when a patient is enrolled in one of our trials, the first thing that happens is that we take a tumor sample and a normal sample from the patient. These samples are then DNA and RNA sequenced, and this sequencing data is then the input that we use in PIONEER™ to design our personalized neoantigen vaccine. The design is then sent forth, and we then have a personalized vaccine manufacturing pipeline that produces the actual vaccine, and then it is then given back to the patient. And in our trials, all our patients receive the personalized cancer vaccine in combination with checkpoint inhibitors, sort of the current state-of-the-art therapy for these patients.
So the PIONEER™ model is the model that we've used to generate several of our oncology product candidates, namely EVX-01, EVX-02, and EVX-03. And in the following slides, I will show you a bit of data from EVX-01, which is our product candidate that is furthest along in clinical development. So it is currently in phase 2, yeah, being tested in a phase 2 clinical trial. But I will show you some data from the phase 1 clinical trial, which, as Andreas said, was performed here in Herlev Hospital. So the EVX-01 is a peptide-based neoantigen vaccine, and it is mixed with an adjuvant, so a molecule that stimulates the immune system, called CAF09b.
In this trial, we only enrolled metastatic melanoma patients, and the primary endpoint, since it was a phase 1 trial, was to see that it was safe for the patients to receive the treatment, but also to see that it was feasible for us to create a personalized vaccine for each patient. The patients received 6 shots, and as I said earlier, they got it in combination with checkpoint inhibitor therapy. So we had some really nice clinical data from this phase 1 trial. First of all, we met the primary endpoint, which was the safety and tolerability. So I don't think any patients had any severe reactions to the neoantigens, which we were very happy to see.
But we also looked into the immune responses that were generated in the patients, and we're very happy to see that we were able to measure neoantigen-specific immune responses in all of the 12 patients. When we look at the clinical data, we saw that eight out of the 12 patients responded to the treatment and two actually had a complete response. So they were basically tumor-free at one point in the study. So we're digging a bit deeper into this data. It was great. But we would also like to see if we could distinguish between the patients that did well and those that didn't. And here we thought that maybe the score that PIONEER assigns each neoantigen that goes into the vaccine would be predictive of how the neoantigen would perform.
So the first thing we did was we saw if the PIONEER™ score correlated with the likelihood of each neoantigen being immunogenic, so triggering an immune response in the patients. And we saw a pretty nice correlation with the neoantigen scoring higher being more likely to generate an immune response. Furthermore, we then looked into how the neoantigen scores of the vaccines that each patient received, how that correlated with their response. And once again, we saw that the patients that received the vaccines, which had, let's say, higher quality neoantigens, so higher PIONEER™ scores, they were also much more likely to get a benefit from the treatment than those that received the, you could say, the low scoring neoantigens. So this is all well and good, but we also decided we were going to look at this at a more unbiased way.
Here, I guess you will recognize this plot from a couple of other presentations. But what we then did was we decided to divide our 12 patients into two groups, so kind of unbiased, those with the highest scoring neoantigens and those with the lowest, and then see how did each of them fare. And as we can see, you can appreciate the blue curve that the six patients that had the highest scoring neoantigens, they had significantly longer progression-free survival. We did the same with the tumor mutational burden, which is sort of the traditional biomarker you use within cancer immunotherapy. And at least in our trial, we saw no difference between those with a high tumor mutational burden and those with a low, indicating that PIONEER is really it does contain some extra information or is maybe specifically relevant for the neoantigen cancer vaccines.
So this data set was really what, you'd say, is the clinical validation of the PIONEER model as well as our AI immunology platform. Now, we've been talking a lot about the EVX-MHC building block and how central it is to our models. So we really wanted to hit that home. So to do this, we designed an experiment where we have our PIONEER model, and instead of using the EVX-MHC building block, we switched that out with two other similar prediction tools for predicting MHC presentation. So one, which is our old gold standard prediction tool, which is based on the NetMHC architecture, so very similar in performance to that, and another highly cited prediction tool called the MixMHCpred. What we saw when we did that was that the precision with which we were able to predict these immunogenic neoantigens, it dropped.
So we were so clearly seeing that EVX-MHC and the newest version of EVX-MHC is critical in order for us to get the optimal performance out of the PIONEER model. Good. So in summary, PIONEER model is our model for designing personalized cancer vaccines based on neoantigens. It's been tested in two phase 1 clinical trials and is being tested in an ongoing phase 2 trial. We showed that the PIONEER scores are predictive of how likely a given neoantigen is to be immunogenic and are also predictive of how likely a patient is to achieve clinical response to the treatment. And finally, we showed that EVX-MHC is a very key building block in the model.
Good. Thank you, Thomas. I have to say, I keep being impressed each time I see those data from the phase 1 trial, even though I've seen it quite a few times. But it is exciting. You had a follow-up question from before regarding the Moderna Merck personalized vaccine. A few thoughts on that without yeah.
Yeah, without taking too much time. Yes. We can always, of course, talk about that for a long time. But I could say, in Evaxion, we are, of course, very excited to see that other companies are succeeding in the field of personalized neoantigen vaccine as they are paving the way for us to also continue. And it shows that the approach has merit. Were there any specific questions about that? As far as I'm aware, they haven't shown anything about how their predictions of neoantigens correlate with response or with the immune system.
No, it was more a comparison of the technology, which was difficult to say too much about.
Yeah. I could say that we don't know a lot about how Moderna chooses their neoantigens. That is a tightly kept secret. But I guess we all here know that Moderna is, of course, an RNA company first and maybe not AI or an immunology company at their core without speaking too much for them, of course.
Good. Then a question here. Are PIONEER quality scores on a 0 to 1 scale? It looks like the top quality score among patients in the phase one study was only around 0.3. Can you help me understand it? And Lee, I'm glad you asked that question because I have been asking the exact same a few times, so.
Yes. Yes, they are from a 0 to 1 scale. But of course, we build in the fact that there is a lot of uncertainty around the predictions. So in practice, I think the highest scoring we've ever seen is around 0.6. But in theory, once we get perfect predictions across the board, it will be from 0 to 1.
And then there is a question, the current status of EVX-01, when phase 2 results will be available? And the answer to that is, I mean, we are in phase 2, and we are expecting the 1-year data readout in Q3 this year. And then another one on the phase 1 study of EVX-03, when will that be started? As we have communicated, then we are not going to start the EVX-03 ourselves, but we are looking to do that in partnerships. And then there's a question here. How will you benefit from a personalized antigen vaccine?
How will you target customers' costs? How will you cope with this potential business model? Pretty broad question, but also super important because, of course, it is a different business model than, yeah, when you just produce an antibody in large scale, and here you have different supply chain, but also a personalized vaccine which is made specifically for the patient and hence a completely different efficacy or potential at least.
Yes. Yeah, I can comment on that. Well, first of all, because we are tailoring the treatment specifically to one patient, we, of course, believe there's potential for a greater effect than a one-size-fits-all approach, right? You could compare it to having a suit tailored for you. It's just going to fit better than one that you buy down in the store off the rack. With regards to cost, that is, of course, going to be a challenge.
But I think we've had some good dialogues with the regulatory authorities and our partners within our personalized vaccine manufacturing pipeline. And there are definitely things that I mean, the risk profile between making a drug for one person and making a batch for 100,000 is very different. So there are definitely things that can be done cheaper and faster in the personalized setting compared to a one-size-fits-all setting. Yeah.
But I think it's also fair to say that, I mean, now we are not looking to bring a personalized vaccine to the market ourselves. And when you get in a pharma partner used to handling large-scale supply chains, I think there are definitely opportunities for optimizing both processes and cost. I think the work that we have been doing in getting the supply chain up and running for our clinical trials, that's quite impressive.
We have actually managed to get it to work very well. So when you talk about larger volumes, more patients, of course, also potential for bringing costs down, right?
Yeah. And there are CAR-T treatments which are personalized that are approved. So it is possible.
Then a good question here as well. Would you prefer further development in a metastatic setting with large tumor burden or probably the easier way in the adjuvant or neoadjuvant setting?
Maybe we should. I don't know if Birgitte is here. This is her. Yeah. Birgitte? Yes. Thomas goes first.
I can definitely answer that. It's a good question. I was at a conference last year where we discussed exactly this. So my personal opinion is that I think the metastatic setting is great because you can show with fewer patients that your treatment is going to have an effect.
In the adjuvant setting, you actually have very few events, so patients that relapse, that you actually need to treat a very large amount or number of patients. And that's fine if you're a big pharma. But I think for us, as a smaller biotech, sort of going in the metastatic, the more challenging setting, that is definitely the way I would go.
And I know Birgitte would have answered the same, I guess, so we don't need.
We actually had a discussion a few weeks ago about the trends in the field. And I think you, yeah, phrased it very nicely. The big companies, they do have more muscles. They do have more resources. And there is a tendency they go for the adjuvant setting where the events are yeah, there are fewer events.
But in theory, easier. Yeah.
Yes. And are there any questions in here? Otherwise, we'll.
Yeah. One last one. Yeah. I don't know how good it is, but I'll try. So I wasn't really clear whether your neoantigens have one mutation or do you look at multiple mutations?
Yeah. Great question. Yes. So each neoantigen generally has one mutation, but we can, of course, include multiple neoantigens from multiple mutations within the vaccine. So in our phase 2 trial, for example, we aim to include 10 neoantigens in each vaccine from 10 distinct mutations. Yes.
And then there is a question here around speed. Whether our platform is faster in identifying new antigens compared to competitors and also turnaround time, you could say. Of course, from having the biopsy to having the vaccine administered matters, right?
Yes. Yes. So yeah, I can say that in our trials, we aim to from when we get the sequencing data to when we have the design, we aim to do that in less than 48 hours every time. And we've basically been able to do that for all patients in our trial. But yeah, I would say speed of design is a very small thing because the production time is actually what takes up a lot of time. And there, I know we have been in our phase one trial as fast as six weeks from tumor biopsy to the patient actually receiving their treatment, which I think is quite fast. But yeah. Yeah.
Then a final question. Immunogenic referred by you refers to humoral response. Would that also work, for example, DNA-based antigenic target that might not have strong IgG response? I'm going to have to read that.
Yes. So actually, not. We are talking about the cellular response, actually, here, not a humoral response. So it is a T-cell-focused vaccine aiming to elicit a T-cell response and not an antibody response.
Cool. I think if no further questions here, then let's say thanks. Well, there's one final one here. Are there databases of neoantigens for defined tumor types where your AI can detect certain public-use neoantigens?
There are definitely databases of neoantigens out there. I will not comment on how complete I think they are. They're probably very far from being so. In theory, we could find a neoantigen that is found in these databases, but we don't use them. Everything is done based on the data we get for the individual patients. Thank you.
I can only say we are looking very much forward to Q3 when we have the 1-year readout of EVX-01 phase 2 trial. That is going to be truly exciting. I think the interim data we showed at the end of last year, super encouraging. And I mean, this is also about the clinical validation of the PIONEER platform, but more important, so making a difference for patients with metastatic melanoma, whereas you say, I mean, even though checkpoint inhibitors are goal therapy, then still very few patients actually do respond, right? So with that, Andreas, will you introduce the next speaker?
Yes. So the next speaker is presenting OBSERVE. And it's Christian Garde, so has been here for ages. And fantastic. And old. So Christian has a PhD in bioinformatics as well and is the key architect on OBSERVE and also on AI-DEEP. Thank you, Christian.
Thank you, Andreas. I'm very happy to have the opportunity to present this work. It's a really prime example of how we can utilize the AI-Immunology™ platform to really facilitate model development to meet medical needs. Just to rewind back a little bit to the previous talk where Thomas presented nicely how we have our PIONEER™ model to design personalized neoantigen vaccines. He also presented that PIONEER™ is able to quantify the quality of the neoantigens. He showed you this figure to the right where you can see that the patients in our clinical trial, EVX-01, those patients that have a high neoantigen quality, they actually have a longer progression-free survival compared to those patients that have a low neoantigen quality predicted by PIONEER™.
But that, of course, also begs the question, so what about the patients who have very few neoantigens of high quality? What can we do about these patients? Can we provide them with a personalized treatment that will be more efficacious than those that the neoantigens can provide? So you really need another antigen source to supplement the neoantigens in order to further optimize the personalized treatment here. And in that regard, I would like to introduce you to the endogenous retroviruses. For the rest of this talk, I will refer to these as ERVs. And the reason I'm introducing you to these would be that they are really promising targets as cancer antigens. What are these ERVs? They are ancient viruses which infected our ancestors thousands of years ago. They have been lying in our DNA, and they have been passed down through generations.
They're very abundant in the genome. Normally, they are not expressed in healthy tissue. Due to the dysregulation that happens in cancers, you suddenly have an expression of these ERVs, ERV antigens. You really have the characteristic of a tumor-specific antigen here. There are a couple of examples in the literature that have described that ERVs can elicit specific T-cell responses in mice, and they can also protect the mice from a tumor challenge. Furthermore, in cancer patients, the T-cell responses specific to ERVs have been measured. Finally, also in the laboratory, T-cells that are specific to the ERV antigens can actually kill tumor cell lines. All these characteristics really hold promise for the ERVs. We then wanted to dive in and make some analyses in order to substantiate whether the ERVs would be a useful source of antigens in the personalized setting.
The first thing we did was to do a very large-scale analysis across a lot of different cancer types on thousands of cancer patients. We analyzed their genomic profile, and we counted up the number of new antigens in the patients' tumors and also the number of ERVs being expressed in the patients' tumors. Interestingly, if you look here at the figure, then you can see that for each cancer type represented each biodot, you can see the median number of neoantigens versus the median number of ERV antigens in the tumors for the patients. And what you can see is that they are not correlated. Furthermore, they're not correlated within each cancer type either.
And that's actually a good thing because then if you have a patient who has very few neoantigens, then there could be the possibility to find an ERV sequence which would be useful for designing a personalized treatment. Yeah. The next thing we did in order to substantiate further whether ERVs as an antigen source would be efficacious, we looked into some cohorts of malignant melanoma patients treated with checkpoint inhibitor therapy. What we did was to analyze genomic profiles from genomic data from the baseline biopsies. We then counted up the number of neoantigens in the patients' tumors and the number of ERVs being expressed in the patients' tumors. Then we split the patients into two groups. A group of patients with a lot of neoantigens get labeled as high TMB. A group of patients with fewer neoantigens get labeled as low TMB.
If you focus on the plot in the center, then we further split the patients into groups based on the number of ERV antigens being expressed in the patient's tumor. You can appreciate that these two curves fall on top of each other. So for these patients that have a lot of neoantigens, it doesn't matter whether you have a lot of ERV antigens being expressed or few. However, on the contrary, if you focus on the plot to the right, then you can see the group of patients with fewer neoantigens. And if you then again split these patients into two groups based on the number of ERV antigens, then you can basically see that those patients that have a lot of ERV antigens being expressed in their tumor survive longer after checkpoint inhibitor therapy compared to those with very few ERVs being expressed.
So this supports that ERVs could be a complementary source of antigens that could help the T-cells to fight the cancer when there's very few quality new antigens present in the tumor. So the next line of evidence we tried to investigate was, well, can we actually find these on the surface of tumor cells? So here we conducted some extra experiments in order to see which kind of MHC ligands do we have displayed on tumor cells. So we did that on two mouse tumors, one called the CT26 and one called the B16-F10. And what we found was that ERVs are actually presented on MHCs. And we also found that our peptide MHC prediction tool, EVX-MHC, is able to predict these. So that's a really good sign that we can actually find these based on genomic data.
So now we had a lot of lines of evidence supporting that we could use this in a personalized setting. So we started developing our model, OBSERVE™, utilizing our AI-Immunology™ platform. And we combined the different elements as such. But if we then look a little bit further under the hood, then I have created a little schematic for you here to really show the flow from patient to the design of a vaccine using OBSERVE™. So the first step is to collect a tumor biopsy from the patient. That is to characterize the patient's HLA type and also identify those ERVs that are specifically expressed in the tumor. Then MHC ligands are predicted, and we rank the ERV antigens, the peptides within the ERV antigens based on the potency of the predicted MHC hotspots.
You can then choose the top-ranking peptides, and you can formulate that as a vaccine, for instance, as peptides or in a nucleotide format, which can then be delivered back into the patient. The nice thing about this is that it fits into an established clinical workflow, and it really aims towards eliciting both the CD8 and CD4 T-cell responses in order to achieve a sustained immune response, but also a sustained effect. Now having developed our OBSERVE model, we wanted to test this out in an animal setting. We designed a personalized ERV-based therapy for a mouse, and then we tried to vaccinate a group of mice with this designed therapy. Then we also vaccinated another group of mice with an empty plasmid so that serves as a negative control.
Then we saw how, then we challenged the mice with the CT26 tumor cell line and saw how well the mice fare. So if you look at the plot to the right, then you can see how the tumor develops over time. It's a time course where you have the number of days after tumor inoculation along the first axis, and then you have the tumor volume on the y-axis. You can see that the negative control increases its volume exponentially as expected, whereas the OBSERVE™-designed vaccine actually completely prevents development of the tumor. The next thing we wanted to showcase was, well, EVX-MHC is a core component in OBSERVE™. So what if we tried to swap this out with a different tool, a gold standard tool? So we tried to do that in order to really pinpoint the power of EVX-MHC.
We did that and also vaccinated mice, and this is shown as the red curve. You can appreciate that this vaccine does not completely prevent the establishment of the tumor. We then also looked into the T-cell responses, and you can see here on the most right figure that the specific T-cell responses towards the EVX- MHC4 actually is 3 times stronger as compared to that of the gold standard group. It really underscores the importance of having a good MHC prediction tool. Finally, we also, more than 70 days later, tried to measure if these T-cell responses were really durable responses. We measured again, and we saw that the T-cell responses were sustained. We also tried to do a secondary challenge on the EVX- MHC4 group to see if they could still prevent establishment of the tumor and was still completely prevented.
So a very strong and sustained effect. Yeah. So now having established proof of concept in the preclinical, we feel that we are ready to really test this out in a clinical setting. So that would basically be in a collaboration with a partner as the EVX-03 personalized trial. And this will be a trial where neoantigens and ERVs will be tested together, and it will be delivered using our in-house developed delivery technology, which is powered with an APC targeting unit. So in this trial, you can imagine that you have some patients which have a lot of neoantigens. You can see that on the left figure all the way on the top. For these patients, they would have a lot of high-quality neoantigens and a lot of the new there would be a lot of neoantigens within the vaccine design.
If you go further down, then you have patients with fewer and fewer high-quality neoantigens. And here, patients would then be given treatments comprising more and more ERV-derived peptides. So it's really a way to ensure that the most optimal personalized treatment is given to the patient. And this is how it fits into the pipeline. So as I said, it would be the next personalized trial this would be tested in. So yeah, we've tested it. So in summary, ERVs, they are a complementary antigen source to the neoantigens. And we have developed the OBSERVE™ model for the design of ERV-based personalized cancer vaccines. And finally, also, we are ready with the paperwork.
We have been in communication with the medical agencies in order to see if this is viable, and we feel ready with the paperwork to really submit to initiate a clinical trial as soon as a suitable collaboration is established with a partner.
Yeah. Thank you so much, Christian. I mean, completely novel target which can be combined in different ways. That is super exciting. And also good to see here the EVX-MHC, how much that actually matters for the predictive capabilities as well. So super interesting. Any questions here or any questions online? We have one here. Are you concerned about dilution effect when combining new antigens and ERVs in the same vaccine?
Well, I believe that this is something that we actually did test, and we saw that the effect on the preventing tumor establishment was preserved. But I don't think that is a high concern.
How personal are ERVs? Are ERV sequences shared or not across patients, and/or do you see differences in ERV expressions across patients? I can tell you so much, we are getting back to that later. Do you want to say a few things without revealing too much of what's going to happen after the break?
Yeah. So in contrast to the neoantigens, the sequence is conserved, and also to a much larger degree is the expression of the ERVs also preserved between cancer patients. So it's definitely a much more viable target for a precision-based approach as compared to the neoantigens.
And we will probably hear more about that after the break. Any other questions? Or we go, for the last break of the day, then we are going to take a break and say, "Oh, how important is the DNA technology in EVX-03?
Yeah. So we had an R&D day last year showcasing how important it is to have this APC targeting unit in the delivery technology. But yes, that could also be moved to an RNA-based technology. So whether it's DNA specifically, I would not comment on the importance of DNA specifically, but at least the APC targeting unit we have seen is a key differentiator in how effective the vaccine is.
Then we have one, "Did vaccine group have any adjuvant?" I'm not perfectly sure what it refers to, but.
No, it did not have any adjuvant, but we did start out with an electroporation in order to get humans to do that.
Cool. Christian, thank you. Let's have a break. So thank you so much for returning, and we are still full house, so that's amazing. I really love it. Now we are jumping into a new session, precision cancer concepts. And there Christian Garde is presenting our AI-DEEP™ model, predicting responders technology.
Thanks, Andreas, for the introduction. Yeah. So now we're shifting gears a little bit, and we're talking about prediction of response to actually an established drug. And so just to start out, checkpoint inhibitors are exploited by cancers in order to evade the immune system. And checkpoint inhibitors, they are an antibody-based therapy that actually impairs this evasive mechanism in order to reinvigorate the immune system so that the immune system can kill cancer. This has really improved the treatment of several solid cancers, and an increasing number of approvals for new cancer types are coming every year. So it is becoming quite widely used and also as first-line therapy. Now, since this has really improved the therapy of several cancer patients, it is also reflected in the market.
Christian already alluded to that the market for checkpoint inhibitors is already now fairly large, but it has actually been projected to reach all the way up to $150 billion by 2030 worldwide. This is a big market. However, despite this being a clear improvement in patient care, then as Thomas also talked about in the PIONEER session, a large fraction of patients actually still do not benefit from the checkpoint inhibitor therapy. That, of course, is an issue. There's really a demand for increasing or further development to increase the efficacy of the immunotherapies. But there's also a demand to identify those patients that do not respond to therapy. That's also namely because as good as this therapy actually is, it also can cause quite severe side effects in some patients. Yes.
To the physician's frustration, however, there are no established biomarkers which are able to predict which patients actually do respond to the checkpoint inhibitor therapy. So that really underlies the issue here. So now I just want you to remember back to the previous sessions where Thomas nicely described how PIONEER is able to quantify the quality of the neoantigens and how that correlates to the progression-free survival of the patients in our in-house combination therapy trial. Furthermore, we also have the OBSERVE biomarker, which we saw in the last session, actually is predictive of the overall survival of melanoma patients receiving checkpoint inhibitor therapy. So we already have established that biomarkers derived from our two models for personalized cancer vaccine development actually can be repurposed to biomarkers predictive of patient response.
So we thought that we are in a pretty good situation here and have sort of a responsibility to develop a new model to help the physicians identify those patients that do not respond to the Checkpoint Inhibitor therapy. So basically, the flow would be that you have a cohort of patients. Some would be responding. Some would not be responding. And you have to identify those that respond. That would be based on a tumor biopsy, which the doctor would collect from the patients, and then a genomic profile of the tumor biopsies, which would then be analyzed by an AI model to discriminate between the responders and non-responders. So basically, this has to be done through the lens, in our view, of genomic biomarkers. And this has several benefits to the patients.
If you can identify a patient that you know will not respond, then the patient can be redirected to an alternative therapy that could have potentially a higher chance of providing the patient with a benefit. Furthermore, the patient would also avoid the risk of severe side effects. And of course, there's also an issue that is talked about now, the impact on the healthcare budgets. So you could alleviate that as well by avoiding to treat patients that do not benefit from the treatment. Yeah. The genomic biomarkers involve private mutations, the environment of the tumor, the immune invasiveness of the tumor, also different sources of antigen burdens, and of course, biomarkers derived from our PIONEER and OBSERVE model. So put in the phrasing of the AI-Immunology platform, we basically take certain elements from the AI-Immunology platform in order to design our predictive model called AI-DEEP.
For developing the model, we then collected a lot of genomic data from 937 cancer patients who received checkpoint inhibitor therapy. The data comprised genomic profiles of biopsies collected prior to the treatment. We then used this data in order to identify the most predictive set of biomarkers, which was then advanced onto the final model development to really see if we can predict patients that do not respond to the checkpoint inhibitor therapy. Yeah. So what we then saw was that if you look at the traditional biomarkers here displayed in the figure with the red curve, then you can see that they actually fail to achieve a very, very high precision, which would be warranted in order to avoid depriving a patient for potentially curing therapy.
On the contrary, you can see that the AI-DEEP model is actually able to achieve the high precision that is wanted, and it does so for 28% of the patients that do not respond to therapy. So all in all, AI-DEEP could actually aid the physician in taking treatment decisions. Yeah. So furthermore, we wanted to investigate. So now we have a range of really predictive biomarkers, but which of them is the most predictive biomarker? So we did that by conducting what is called a feature equation study within the field of machine learning. And that is basically a study where you estimate the importance of each of the biomarkers on the predictive performance.
What we then discovered was that our in-house developed biomarkers derived from the PIONEER™ model and the OBSERVE™ model actually ranked among the most informative biomarkers, whereas the more established biomarkers, the tumor mutational burden and the PD-L1 expression, actually ranked as being much less informative. Yeah. So in summary, we just want to list here that AI-DEEP™ leverages our PIONEER™ and OBSERVE™ model in order to predict whether patients will respond to checkpoint inhibitor therapy. And it can do so quite accurately for a subset of the non-responding patients. And we are currently exploring opportunities for commercial offering and also further clinical validation as a companion diagnostic. Thank you.
Thank you, Christian. Super good presentation. But try to think about it, right? You are diagnosed with cancer. It would be nice to know upfront whether the therapy is going to work or not so you don't have to spend three months on a therapy with a lot of side effects, which then turns out not to work. That's one thing. Another thing is $150 billion by 2030. This is a huge market. If we can just support a little bit more efficient use of checkpoint inhibitors, it would mean a lot for societies and healthcare burden. So I think this is one of the areas where, even though it's slightly different than some of the other things we're doing, we have the obligation to do it because it is repurposing things. It's taking some building blocks, putting it together in a different way.
Then it is about the data and predictive capabilities we have. It might seem different, but it's in essence in the core of what we do with AI immunology. This is going to make a huge difference for society if we are going to be successful in developing a companion diagnostic or whatever it's going to be when we get to a commercial offering. Questions?
Have you tried to use data on microbiome, so bacteria, to, let's say, enhance the predictability of your AI-DEEP model?
No, we have not tried that yet.
Would that be one route?
That could be a way to expand the set of biomarkers. Furthermore, that's true. And there could also be a different scope or set of biomarkers, for instance, derived from scans or others that could also complement the genomic biomarkers. Yeah.
And then we have a question here. Some PD-1 inhibitors are approved in patients with certain PD-1 levels. Are you saying that your tool could be more exact as a response predictor?
Yes.
That's exactly what's so exciting here that this is going to be more precise. We have shown it's more precise than what's available today. Any more questions out there in the online group? Otherwise, we'll say thank you, Christian. Then it's almost a little bit sad that we're coming up to the final presentation, but it's also going to be super exciting. Do you want to introduce the next speaker?
Yes, definitely. So Jens Kringelum. So PhD in bioinformatics, VP of AI and innovation in Evaxion, the first employee in Evaxion. And Jens is going to present some very, very exciting avenues we can take with OBSERVE.
Yeah. Thank you, Andreas. And thank you for the introduction. So first and foremost, I'm very proud and honored that I'm allowed to be the last one. Thank you. But I also know that I'm the one standing between drinks and snack. But I hope we still have some energy left for this because I think this is actually truly exciting. So we tried to keep a little bit of mystery in the title of this talk. So I was a little bit sad when the very good question to OBSERVE came, took away a bit of the thunder. But I can tell you already now that, yes, you can make shared and precision vaccine from these endogenous retrovirus targets. And this is a little bit what I've been showing today, how we do that in Evaxion.
But first and foremost, let's just remind ourselves on why this is actually important. So this figure here, this is probably the most famous figure in cancer research. It was first published in 2013 by a large consortium of global researchers. And what it shows is that it shows how many mutations are there in different types of cancers. And what we see to the right, that is the type of cancers that has a lot of mutations. And then we also see that to the left, there are numbers of cancer types that have very few mutations. And as Thomas was explaining, these mutations are the source of modern personalized cancer vaccines. These are the input that we need for PIONEER to find very good neoantigen sources. And as Thomas also showed, this is important for efficacy.
So even though we have a great system, there might be some cancer types that are just very difficult to treat because they have few targets for these modern cancer vaccines. So luckily, Christian, there you are. Luckily, you came along and showed us that there are these endogenous retroviruses in the human genome. So these are the ones that have been inherited from our ancestors, and we all have them. And they constitute a huge part of the genome. And normally, they're all silent. But as Christian also showed, these, for some reasons, start to be used in cancer cells. They're not used in our healthy cells, but they're used in cancer cells. And these are actually extremely good targets. So after realizing this, we start to look into, so what does the research literature actually tell us about this?
Is there any proof that this works in patients? There are. If you're looking into AML studies, we see here that people have actually done a lot of great research on this. What we see here on the left is a research group that in 2020 published some data showing that just from unvaccinated patients, they could see that the immune system of these patients actually picks up some of so they find that the immune system already, without being primed by a vaccine, is already picking up on these endogenous retroviruses targets. What we see in the middle of this is basically the same story, but from a different research group, that they can actually take out blood from patients and stimulate that with these specific targets, and they see an immune response. This is quite remarkable.
It tells us that we might be onto something that can be used in a vaccine. What we see here to the right is that these are actually displayed on the surface of the cancer cells, just as we saw in our experiments in the mouse cancer cells that Christian was talking about. We also see that when we look into humans. So these are important data for us to make decisions on going forward with these kinds of targets. The next thing is, it's all about data. One thing we have done recently is to gather a dataset of more than 15,000 patients and then run them through our PIONEER model and also run it through our OBSERVE model.
When doing that, and this is just for selected cancer types, when doing that, we see that, just as Christian was explaining to us, that for some cancer types, we see a lot of ERVs, but that is not correlated with the number of mutations. Here, I have just highlighted a specific class and type of cancers, and that is blood cancers. We see that for some of these blood cancers, they just have a lot of these endogenous retroviruses. Of course, these could be very interesting cancer types for us to explore further. Now we are going into why this is so exciting. One thing we also realized was that some of these targets are shared among patients. That is very rare when we talk neoantigens and we talk mutations. But when we talk about ERVs, this is actually a common phenomenon.
Another realization is that these are just viruses. They look a bit different from what we are used to. But back in the day, when they infected one of our ancestors, they were just viruses that were then integrated into our genome. We have already built an AI immunology model that is very effective in designing virus vaccines, our RAVEN model. So why don't we take building blocks from the RAVEN model and build that into our OBSERVE model? That was exactly what we did. So we took the building blocks that are needed to build shared and precision vaccines from the RAVEN model and built that into our OBSERVE model. And thereby, we created what we called the OBSERVE 2.0 AI immunology model.
By doing this, we were then able not only to do personalized T and neoantigen vaccines, but we were also able to make shared vaccines where you make one or a couple of vaccines for the entire populations of cancer patients within a given type of cancers. We took the new OBSERVE 2.0 and just employed it to a number of promising cancer types to see, is this actually possible? This is the results. Basically, the answer is, yes, it's possible. But I will also tell you what this figure means. It is maybe a little bit complicated. So what we have here on the x-axis, that is number of epitopes, a number of targets that the given vaccine would target each individual patient.
What we have on the y-axis is the likelihood of the vaccine actually inducing an immune response in any given patient in that population that is having this type of cancer that will then be able to target this target in the cancer cell. So what we see is that if we are just requiring the vaccine to have 1 target, if that is enough, that is normally not enough. Then what we see, for instance, in ALL cancers and AML cancers is that we, with a very high likelihood, just with 1 vaccine, would be able to induce an immune response in that patient that will target at least 1 target. And then we have 2 targets, 3 targets, 4 targets, 5 targets, 6 targets, 7 targets.
If you look at 8 targets, we see that the likelihood is still very close to 1, meaning that it's very close to 100%. If you then go, say, if you want 25 targets, then it goes a bit down. But what we also see is that, for instance, for lung cancer, it's very difficult to make shared vaccines. So here, the concept is much more difficult, which means that in that case, we probably need personalized vaccines or precision vaccines. Another thing, we started to look into these different cancer indications. One specific indication we looked into was AML. This is a broad cancer that, unfortunately, hits children but also adults. When we start looking into the data that we have, we had one dataset for children, and we had another dataset for adults.
We wanted to see, so do they have the same targets, or do they have different targets? When we start looking into that, the first thing we did was to look, do they have the same number of F expressed? What we see is actually that the children, if you are a child below 20 years old, are being diagnosed with AML. You have a much lower number of F expressed than if you are an adult above 20 years old. These could be two different diseases, actually. What we investigated was whether this expression of F correlated with disease states, so how severe the disease are. That was not the case. It simply only correlated with age. There are some interesting aspects in that that we wanted to explore even further.
So we looked into what F was actually expressed in adults and in children. What we saw was that it was basically the same F, but adults just expressed more. So they have all that is expressed. Then if you are an adult and diagnosed with AML, you simply just have more F expressed. So we took our OBSERVE 2.0 model and said, "Okay, but could we use this information to make even better vaccines?" Then what we did was that, why don't we do one vaccine for adults based on the F expression in adults and another one for children? What we see is that if you do that for the adult population, it doesn't make a lot different because they are already having all the F expressed that we were targeting with the fully shared vaccines.
If you look into children and just look in the case where we want 25 targets, here, they are a huge difference depending on whether you make shared vaccines for all AML patients or a precision vaccine based on which F is actually expressed in these different patient populations. And this is, of course, something that we are going to explore even further in other indications. The split doesn't have to be in children and adults. It could be based on co-expression of F, based on HLA types, so what kind of different immune system do we have? Right now, we are exploring how we can use this platform even further to make new pipeline products. My job currently is to explore how can we make this as fast as possible into the preclinical stage and, of course, also later on into humans.
And with that, I hope that I convinced you that, again, EVX-03 is a promising target, especially for these cancer types that are difficult to treat with our current modern vaccines. Also, we have developed this OBSERVE™ 2.0 model that has building blocks from RAVEN™, which allows us to do these shared or precision vaccines. And of course, these are great new avenues. I think we had a question before on the manufacturing costs on personalized medicine. Of course, if you have a shared approach, the approach to market is different from when you have a personalized approach. So with that, I hope that I have just sparked your interest a little bit in this and just maybe open up the window into what is coming next in the field of cancer vaccines. With that, thank you.
Well, I think, Jens, the good thing is we know how fast we need to do it because we have promised that we will have a preclinical proof of concept for our precision-based vaccines in the second half. And this is exactly why I'm looking so much forward to that because it's a truly exciting concept combining, you can say, some of the great things of personalized vaccine with some of the simplicity of off-the-shelf solutions. So that is very exciting.
Exactly.
We have a question here around, can orphan diseases, not necessarily cancer, be a target for Evaxion considering that fewer patients need to go to clinical trial and fewer clinical successes can trigger approval for treatment? You can say, of course, some of these cancers are qualified as orphan diseases, but it all depends on our model and how it fits into the model. I don't know if you have a few thousand.
Yeah. I mean, I think that's definitely right. I mean, often diseases offer us a lot of great possibilities. And also, our models and building blocks can be applied in other disease models, as Andreas was talking about. And in this case, there are also other non-cancer diseases where these endogenous retroviruses play a role.
Jens.
Oh, that's true. But it's also fair to say we still have ample opportunities within cancer and infectious diseases. I know we talk about additional therapeutic areas, and that's also exciting, but we still have ample opportunities where we are. But we also owe it to look beyond where we are now given the capabilities we have and the platform we have. Then two questions here. Can you please clarify if OBSERVE 2.0 enables designing precision or shared vaccines? That's one. Have you looked into FLT3 expression patterns among AML patients with various genetic mutations such as TP53 and IDH?
Yeah. I can tell you the last question. No, we have not looked into that yet, but that's on our agenda. The second one, so it basically does both. So how we envision going from one fits all to fully personalized, that's sort of a sliding window. So you can say that for some specific indications, some specific diseases, there, you need a fully personalized approach because that's just the nature of the disease. In other scenarios, you can, for instance, like in ALL or other blood cancers, you can have one fits all approach, and then you can have everything in between there. So basically, what our system can do is that if you tell that you have data for an interesting indication and you want to make a vaccine for that, it basically tells you, can you make a shared one fits all?
Do you need to go to a precision approach where you may need 2, 3, 4, 5, or do you need to go to a fully personalized approach? So it can give you that number, basically, how many different designs, different vaccines do you need to make to cover this population and this disease and hopefully have an effect in patients?
We have another one. Does a vaccine use mutations in non-coding regions to predict new epitopes?
That's maybe more for you, Thomas .
Yeah. No, we don't.
Yeah. No, we don't.
We are actually fine. No, we don't. But also, I mean, we can say that at least for the personalized approach, we do depend on mRNA sequencing data. So we do know whether it's there in the cancer cell.
No questions in here. Then I think it's thank you, Jens. Super interesting and definitely looking forward to. Then I guess you know what they say. All good things come to an end. That's also the case here. But I have to say, I do hope that you understand why Andreas has been looking forward to this for 15 years because it is a highly unique platform. And it's a platform which has been perfected by a very strong team. It's a platform which offers immense potential for addressing unmet needs, which is in the end why we are here. And it's also a platform which has a lot of future perspectives. So it's really been a pleasure having the opportunity of going through some of the details here.
And I mean, without repeating things, of course, I do think this is clearly a new era in discovery, design, and development of vaccines. I think the comparison up against reverse vaccinology, two companies spending we don't know how many years, but many years on coming up with those eight possible targets, even doing it in a very quick way, that's super exciting. And then, I mean, just take the precision concept here as well. So many examples of what is a new area and new things. It's all about, yeah, the enablement with AI. I'm not going to talk about that. There's too much talk about that. I think we prefer rather just doing it than talking about it and then having the outcome matter. And again, this is about addressing the unmet needs.
It's also the first time we've given the insights into this modular architecture of the platform because often had this question, why are you doing so many different things? Does it make sense? Well, we are actually not doing many different things. We're just putting things together in a way where we can do a lot of difference or make a lot of difference. And in the end, that's what matters. And then, of course, the fact that we have our platform validated by partnerships already. We have a strong focus on generating partnerships because in the end, bringing these offerings to patients that do require partners, we are not going to be a commercial-scale organization for the foreseeable future, if ever, but we are going to partner, and that's what we have a strong focus on. And I'll, of course, always welcome partners to engage in dialogue.
We have exciting assets, and I'm looking forward to bringing that forward. So, Andreas, do you want to come on stage for a few conclusive remarks? Was this worth waiting 15 years for?
Yes, it was. Yes, it was.
Yeah, exactly. So I'm truly grateful of this. And also, I mean, I feel like that we are very close of just like a rocket up in the air. So I think really, this is really, really interesting. And let's see in one year where we are. But I hope you'll come back and hear more.
One important thing. Tomorrow, all of the participants, also, you are there online. You will get a mail from us with a very brief questionnaire. We would very much welcome your feedback as to what worked, what didn't work, so we can do even better next time. We like to continuously do better, and that's where we need your feedback. And then I just want to say, first of all, thank you to all the presenters and not least the comps team who have been doing a lot of work and getting us ready here. And then thank you to all of you for showing up here. It's great to have guests in the house. And those online, I wish we could invite you for the drinks in the kitchen area here, but we'll do that next time. And thank you so much to everyone for joining.
It has been a true pleasure having a lot of good questions, and we'll follow up on those. We didn't get to answer yet. With that, thank you so much, Jordan, for a very active discussion and good dialogue.