Applying AI-Driven Biosimulation to Drug Discovery

Ah, okay. Thank you. It’s a super advanced device here. That okay. That fits in with the conversation about biohumidation.

We have a good advanced devices. Okay. So being here today, I think there’s kind of like three objectives in the Here we go.

In the conversation here today. First and foremost is a little conversation about, well, this whole idea of biosimulation that was kind of like more born later stages in drug development, is it actually meaningful to think of this in the context of discovery? And what is the status actually of biosimulation more kind of like as we go into the clinical domains? And has it proven that it can actually deliver the results?

And if we then think about that and we deliver on the context of that, well, if that is then really proven, well, and what would it be meaningful to then apply it within a discovery environment?

Then last but not least, okay, if we assume that all of those things will then happen and we’ll have success with it, what does that mean in discovery in itself? What is it we’re going to do? What is it the opportunities are? And what could the impact of those models be?

So let’s discuss that a little bit. First and foremost, around the actual adoption.

So biosimulation has been a topic forever. Right. It has been around us ever since the early 90s in different shapes and forms. But it was not before kind of like two thousand and five ish that the first documentation of the results of bio simulation really began to come up.

Down here in the lower corner, there was an article that was published actually originally by Pfizer, by their model informed drug developers to talk a little bit about or the modelers if we should just call them that to talk a little bit about what the impact of using model informed drug development would be and the impact would be. And they went back and they analysed the impact of the different drug programs they’ve had. And the result was fairly impressive. Right?

There were savings across the board both monetarily speaking, resource wise and time wise. So okay, that’s where we started.

The next question you then ask yourself say, okay, there’s something about it. Over the years, this has grown. This has become more and more and more. Well, what does it actually do? And what does it focus on? And what could that connection be then to discovery?

Well on on the left hand you see that it focuses kind of like on the PK side and on the right hand side here the PD side focuses on what does the drug actually what does the body do to the drug and vice versa what does the drug do to the body. And before you even get to that you naturally need to understand what drug are you putting into the body, in what dose and how does that actually react. Does it have impact on formulation and everything else? And that holistic picture is what model informed drug development at the end of the day is all about.

So since back then, it has evolved a lot. Seterra is one of those companies, that’s where I’m from, that are delivering applications in this space today. We’re not the only one.

On one hand on the PK side, we deliver a set of simulators that basically do PPPK modeling of different kinds and help you predict. Typically when you sit from candidate selection to phase 2A and really optimize around the parameters and the biosimulation that is necessary there. Helps you decide upon whether you should stop the work you’re doing or accelerate it and it can even help you avoid doing actual trials both on the animal side and on the human side as well. And that is also helpful when you have drugs that are already approved where you want to extend the label for the drug itself. The other half of it is far more complicated and that’s the part of what the drug does actually does to the body.

Here you actually need to simulate the entire biological system, right? You need to understand what is the biological pathways the doc operate within, what is the biochemistry it operates within and how can you actually impact that with that analysis. So naturally there’s a lot of sub variables here that you need to understand in system biology in itself, in the way that you actually think about biomarkers and the other indications of your drug. And we have developed applications to help with that as well. Now, the key characteristic that we have talked about in all of these presentations today is this does not really work without a scientist.

Right? This is not an application where you now can put a normal human being, if you will, that is not a scientist behind a wheel and then we just start striving. This works hand in hand with highly qualified individuals that can actually drive and understand the context of these applications that already today are loaded with, you know, AI capabilities end to end in many of the offerings that you can find in the market today.

Okay. So it’s validated more than twenty years ago. There is applications out there today. How is the regulators viewing this and what’s the adoption of it in those clinical stages? Well, the adoption is very high and the regulators completely adopt it and they actually accept these submissions today. They subscribe to these submissions. And there is actually guidance, for example, for the Seterra technology offerings from FDA or EMA or PMDA on how to actually do your submissions and how to operate with these software applications in the context of the submissions itself.

Maybe a little proud moment, you know, eighty percent of all drug submissions you can see up there in the round corner comes from our sim sim simulator in the PPPK area.

QSP is newer, right? You know, so whereas PPPK and that was kind of like the Pfizer origin and the thinking behind it. QSP is kind of really taken off as biological drugs has taken off. And the models and the complexity is getting higher and higher where you have combination therapies today.

You have really high standard of care. You actually need to meet to put new drugs into the market. For many diseases and so forth. You do need these QSP models.

And what is also good to see is that there is actually a strong adoption on that. You can see it’s not that many years ago, twenty thirteen, where you can barely see the bar, but it has grown. And it is growing at a phenomenal rate today. And I think most large biopharmaceutical companies hire staff today within this domain and biotech companies are using external consultants to help address these topics that they can’t hire the staff for themselves.

So all in all that is good as well.

And here you have a set of drugs where they have actually been used to submit an extension of the label. It could be pediatrics, could be any other kind of extension of the label for that specific drug that a clinical trial was avoided. And if you in one case avoided a phase three clinical trial which more than thirty percent of these represent, is naturally a lot of money you will save.

Okay, so we have established the foundation.

There is a bio simulation world out there. It is very, very well founded with technologies today. They are available. They can help you do these studies in a clinical environment, and they are adopted by the regulators. They are widely understood, and they are being staffed in all biopharmaceuticals around the world.

Next question then is, okay, is it meaningful to take this into drug discovery? What what can it actually help us with?

And you know, think a good way to look at it is to look at phase one failures.

Where are we when we begin to put a drug into a clinical trial and what’s the failure rate? And if you just look at some of the drugs themselves, the failure rates are very very high.

So this actually means that we enter into a clinical environment not really knowing a lot.

Still gathering a lot of scientific insight and they will naturally never come a situation where we can take all of that predictability and put it all up in a discovery environment. But we clearly can do a lot better than we’re doing today. What we can see is that for modality types and for disease areas where there is a lot of drugs in the market, knowledge and understanding is slightly better. But it’s not much better. And you can see it the other way around if you instead you go in and look at modalities, you have exactly the same picture.

So okay, there is something to be done. There is an objective here of how can we take some of that biosimulation knowledge that is gathered, you know, kind of like from a candidate selection state as you begin to move into animal testing or any other type of new approaches or NAMS today into clinical trials and how can we face some of that knowledge and impact decision making and discovery in a way that we don’t do today. It’s obvious that that is the answer. And the answer is we need to take those two modeling approaches that I just discussed and blend it with the entire PKPD knowledge domain and then execute upon that as we just laid out.

So the question then is, has anybody done that? Has anybody thought about that? Do we have any use cases today that can kind of like guide us in that world? And there’s a couple of examples here that I’ve joined just joined up.

Genentech did a interesting study back some three-four years ago where they tried to go back and look at the PK profile for eighteen drugs and see what could they have predicted using those PK prediction at an earlier stage of the drug and would fundamentally have changed how they had looked at them and how they had executed on them. And if you go across more of these drug domains, yes, It’s okay to continue. I’ll check on that. Yeah.

Yeah. It’s all good. We’ll check it out. Okay. The same happened at Sanofi.

Sanofi is one of the really advanced companies.

They have implemented more than a hundred different AI algorithms into their discovery domain. One of them is to predict rat PK parameters already in the early stages. This is specifically helpful if you want to discard potential candidates very fast. And we’re helping, are working with them as well as others on that and that is very successful. And there’s two other examples here where AstraZeneca has done the same and where we are taking it further.

The last thing then naturally is to figure out okay, so first and foremost we all talked about how it impacts biosimulation the clinical trials. It’s proven, it’s adopted, it’s there today, we just need to go execute. The other one is what is it we need to do to take it into the discovery and is it worth doing? The data is there, shows that it’s worth doing, we should go do it and we should figure out how to take these models further up.

Okay. So let’s try to say, okay, how could we now actually go execute this? And, you know, naturally, please excuse me for not being all encompassing here. I’m using our examples for Zetera, to use our software products and what else we could do here.

So when we sit in discovery, our main dashboard, if you will, or our main control cockpit is an application called D three sixty that helps you select these molecules and helps select which one is actually viable candidates to move forward. And the question naturally then is those two applications that I illustrated before on the PKPB side or the PK side and our IQ simulator on the PD side, would it be possible to take knowledge and data from that and actually derive that into earlier decision making? And the answer is yes.

We have experimented with, you know, I won’t write all of these up right from the log P, log S and all the other ones. But all of these variables, are beginning to gain the knowledge and understanding of how to drive that up into a discovery decision making phase. And it’s already also implemented by for example Sanofi that I mentioned a little bit earlier.

I think another good example is to be more specific.

So DELI is a very drug induced liver injury is a very interesting topic.

About one third of all drugs fails because of this, and they fail very often when you get into clinical trials. So you are actually in the human and we are talking about that PD thing. You know what the drug actually does to the body and the knowledge you can get. And the understanding of the impact of biomarkers and other things that actually drives that daily.

And what is surprising is that by doing these simulation and taking these data, you are actually capable with an almost one hundred percent success rate to predict daily in discovery.

So that in itself means that we’re in a situation where throughput if we really execute and embed this within the biopharmaceutical and biotech companies in general, we should be able to see a thirty percent increase in ability to actually optimize better and derive better results. Another one is on secondary intelligence.

So this is more kind of around unintended effects of the drug, things you didn’t expect.

Right? And and those of you in the room that are QSP modelers today know that this is the task of QSP modeling. But this is kind of like finding out the pathway of the biomarkers to really understand those unintended effect and how can you actually talk about them in the drug itself. And we also understand how to do that.

And the application we have built is kind of like not a precise science with kind of like a curve. It’s more like a traffic like signal, where we indicate kind of like is it red, yellow, green here? What do we see? And how do we operate on this?

So there is example of workflows. There is things we can do besides deriving all of those other indicators that I talked about before that is fundamental for us to actually execute upon.

The third thing is more degenerative AI piece. Right? So all the AI components that has been involved in this conversation so far has more been about ability to leverage data, optimize the data, search the data, categorize the data, be precise about the data, and then do predictions of different things, which is a blend of machine learning and AI today in its best form.

The last piece is naturally, you know, what you have down there at the bottom is to generate novel compounds. And I think there’s different categories of areas that have high interest.

Right. And and what we see in the conversations that we have is that there’s a very, very strong interest that when you have a domain of potential molecules you’re operating with and you you say, I I’m I’m looking at this molecule range at this optimization. How can we use generative AI to actually optimize that? So not looking at the endless universe of potential molecules to be generated, but looking at a very very narrow scope to really truly optimize what it is we’re already focused on. And specifically with the complexity we have today of drug categories, right, I talked about them before where it’s you know multiple different things you need to handle at once. This is a good example.

It is very hard, right? I still think that you see a lot of examples out there today of companies that are in this generative space that go into their own lab testing and their own generation of these molecules. We will probably be better doing it ten years from now, but potentially narrowing down the domain and narrowing how we think about it could be more efficient and actually gain us more results.

I think one of the things that is very important in the whole picture is to think holistically about it. So that’s what we have tried to do here at Zatera. Again, I understand there’s other potential solutions out there today, but the thinking of how does this entire scientific innovation life cycle actually work and how do you encompass the different things that happens in that? And what is it you need as you get to those different stages?

Right, I think that there’s been a lot of other conversations are very overheard here today is kind of like, well, we need all the data to be better to run AI on them. Then other people are talking about the need of speed and so forth and so forth. I think there’s other variables that are very important as you think about this. The preciseness of the science, right?

If you think about the preciseness of the feedback when you sit in a a in a phase three clinical trial, the data needs to be incredibly precise.

And there’s very little room for variability. There’s very little room for error. And you need to collect data, you need to analyze data, you need to understand everything accordingly to that. Whereas if you sit in a discovery space, that preciseness is not needed to the same level.

The speed is much more important. And the velocity of the decision making and kind of like, yeah, we’re good enough on the data we know there and that accuracy increases. Right. So I think it’s important to understand that in that context and operate with the bio simulation modeling and how you use your underlying AI for this.

So that that’s all good. That’s all the models and and everything else you need. But there is actually a topic that’s way more important.

One of the things that characterizes drug discovery today is this very siloed thinking and very siloed execution there is, generally speaking in biopharmaceutical company and biotech companies.

We tend to do the modeling in one team and then we have some other analysis that’s done in another team and blah blah blah. You know the whole story. The collaborative framework and how we can actually make this a holistic topic where the people that are involved in the project can share better. How can we share the knowledge? How can we share the analysis? How can we share the decision making are naturally fundamental. But what is also fundamental is how can we share the data?

Right? Do I have data that I’ve used in my analysis to do my decision making that can be used in a net next step and a next analysis? And what do we actually do to make that entire data sharing possible and that execution within this context possible as well?

So that’s one of the things that we are very focused on. I won’t say we have a very beautiful solution today at Zatera on this topic. I actually don’t think anybody has a beautiful solution on it. But I think this is one of the things that over the next couple of years, as we see agentic frameworks grow up, there will be opportunity for this. I actually think there will be opportunity for then blending all of these scientific disciplines together and begin to see this execute. So I’m excited about that specific element myself.

So all of that naturally in itself then has to kind of like categorize and group itself into some type of platform thinking. Right? What does that platform look like? And and what is it that’s needed for that platform to really operate? And we believe that, you know, it’s kind of like these four kind of levels of execution that’s really needed. First is naturally the foundational models that you need to be able to understand whether they are they are empirical models or they are you know mechanistic models or other type of models and how do you operate the datas on top of that. So that all needs to be understood.

Then you need to understand the context of the workflow it operates within and what are you actually trying to do and how are you trying to execute. Right. So again, if you sit in an area where you’re trying to understand specific execution of specific modalities of coupling the molecules together, need to be able to orchestrate that. And then on top of that comes all of these agents and I’ll come back to that in a second because that is super important.

I think the way we think about this today is we think of it as human beings in general. We think about, you know, over here is Peter, and then there’s Alice that works together with Paul. And we have these three people working together, and how do we make the data flow between them and the decisions they make and everything else?

The the thinking we need to begin to adopt, right, is we need to begin to think of these agentic agents that we can put up and this AI that we can build around them as an army of these scientists that can kind of like accompany my me and my work.

And that can be faces between me and other real human beings and or between the tasks I need to do. And I think we have not orchestrated that that well yet in the scientific world. I think in non scientific disciplines it’s actually going a lot better. If you look at kind of like what Copilot can do today in the office environment and other tools like that, not bad. Right? But when we talk about the orchestrations of these real true scientific disciplines where there is a very deep meaning between the data and the impact on the workflows and the outcomes of specific results, we don’t have that in place today and these barriers are clearly there and needs to be overcome.

So all in all my my final comments here.

The first one is, you know, kind of like self explanatory. Of course, we believe in this. This is the whole reason why we’re embedded on this journey. I think the second point is kind of like super duper important and this is what I measured a little bit and what I talked a little bit about before, right? How important is what type of accuracy, what type of speed, what type of detail, at what point in time to make the right decision.

And I think you all of you that works in biopharmaceutical companies today know the situation that you know, hey, I I need to do an early feasibility assessment on this potential candidate, but I’m I’m not really sure how to go at it. And I know there’s this consultant out there in the world that I could go and hire, But at my company, the procurement process to go and get the forty hours of work I need from that smart person, I can’t do it because the time to get that all of that contractual work done is too long. And I actually need to make a decision in a week about this specific thing. I can’t wait.

Right? So the whole facilitation of making this happen and the orchestration of how we actually do that work is super duper important. And then the last comment here, which I I believe I’ve addressed well. Right? We need to tear down these silos.

And I still think even though we respect each other as teams in the discovery area and there is high regards for all the scientific knowledge and everything else, we can work better together. And I think these platforms actually need to help us achieve that.

So that’s what I wanted to say today. Thank you.