0:00
Thanks a lot.
0:00
Thanks a lot for being so curious and showing up because you might have heard nothing about SynSilico so far and a lot about AI.
0:08
What I tried to do is explaining what SynSilico does actually to make AI or pharma.
0:20
I think we all share this belief that AI is fantastic and it's changing the world.
0:27
I usually start this presentation with a ChatGPT search where I ask can you give me a quotation what AI will do in the future?
0:37
And ChatGPT tells me Lynn Fromme is proposing that.
0:42
And then I'm looking at Lynn, who is Lynn? Lynn doesn't exist here.
0:46
And we have to sometimes face those issues that AI propose something that doesn't tell me.
0:53
That's why I also want to say that SynSilico is not only about AI, but it's actually also about the humans and understanding of biological systems that matter for us.
1:04
It is a joint venture.
1:06
The one side of the company brought AI capabilities, conversational capabilities together with the other part, which seems to be the DSM pharma development unit industry.
1:18
Now together AI and life science capabilities.
1:22
We have a small team at the start.
1:27
The most important people are the people that deal with the computers.
1:31
Because I'm a biotech guy, I have no idea about the computer systems and the codes and so on.
1:38
But I can pretty easily tell that what the machines, it's all this solution does not really work because angles are wrong, and the binding modes are wrong, and the dynamics are not taken into account.
1:50
So that is how we operate there with within our company there are predictions and there are rational people behind that can tell us if those predictions are right or wrong.
2:01
What we have put together initially was our mother company InnoSyn, they used to develop pharma products or the processes for pharma products for pharma customers.
2:14
That was our starting point and for such activities Bayesian optimization helps to optimise processes.
2:23
They it helps to optimise actually anything, even bread baking and so on.
2:29
That's why we initially started to implement Bayesian optimization protocols to help our chemists to improve their experiments.
2:41
Nowadays those tools are also used as APIs for developing self-driving [unclear].
2:47
But there is something we can discuss a little bit later.
2:51
The other part was in the insulin discovery field where we took advantage of AlphaFold and other models to rapidly screen binding modes of substrates or products to proteins.
3:07
But if you're biotech experts, you exactly know that binding modes don't predict catalysis, or we need to do something else.
3:17
And that's why we've introduced the strong molecular dynamics capabilities in our companies to go a couple of steps forward and just predicting binding modes because that doesn't help.
3:29
On the Bayesian optimization too, I think that I’ll jump over this.
3:37
You can find a better version of our tool on the Internet.
3:41
You can simply go to optimizer.synsilico.com and then you can provide it with your experiments.
3:47
You can design the input parameters and a lot of different things to be more efficient with your experiments.
3:57
It's building up also a database.
4:01
It can be nicely used if you have more and more data set up.
4:07
You can use this database to train models next to our optimizer proposing as a system.
4:15
The enzyme discovery toolbox, protein folding, active site identification, and docking.
4:24
And we could actually identify nicely enzymes with this toolbox.
4:30
But now we want to go beyond that and more qualitative in terms of activity results. And people say they can predict activity from sequences, they do, they [unclear] data, but if you go outside the data, we at least find [inaudible].
4:52
For the pharma relevant inhibitor design.
4:57
We also have this pipeline we go into the structure, binding site, dockings, AutoDock Vina, this is a tool everyone is actually using.
5:07
But the predictions that we get here are insufficient for a high quality hit.
5:14
This is at least our view and therefore [inaudible] they know how to design drugs. And we need [inaudible] to design.
5:33
With these initial workflows, we had also the problem that if we were proposing a hit, yeah, our chemists or our customers’ chemists required the lens actually to synthesise those molecules to define or to give us a response.
5:47
Is it a hit or not? Yeah.
5:49
And then that is still valid.
5:52
You can do this.
5:53
But we wanted to do some or do something else and benefit from our protein engineering capabilities.
5:59
And for that reason we jumped now on a workflow that is designing single chain antibodies.
6:07
Also there, people know how to do this.
6:09
There is a generative AI that can identify binding molecules.
6:16
But also there you have poses wrong, you have a lot of things wrong with diffusion models and therefore you again have to go into specific designs that is empty support.
6:27
And then you can read our physical data with such workflows.
6:33
Nice thing about this platform is as I said, it's fast.
6:37
We can not only design it, but we can also produce the antibodies rapidly with microbial systems and can directly ship them or do the assay in house.
6:47
And this is changing, I think, the world of how you then can identify targets or you can actually also look at the pharmaceutical products, but also you can select if you fish out a protein without his text also and then purify proteins out of mixture of our proteins.
7:12
And it's of you're using this platform.
7:16
And I have a couple of more minutes.
7:20
If you know camelid-based antibodies, it's single chain, it's small.
7:30
If the features that you can recognise a lot of different epitopes.
7:37
This is one of the reasons why we've chosen that platform and they are clinically proven and economic to be produced.
7:45
And this is something, the perfect starting point for such a platform.
7:50
And you know Ablynx and Sanofi have spent quite a lot of money to prove that this is feasible.
7:58
We built now a little bit based on what it did, but of course without infringing any patents, [inaudible].
8:05
But we can apply those screens.
8:10
Yeah, just to emphasise that it's not only AI, it's rational design.
8:15
It means that you need to apply those workflows.
8:20
Otherwise, you know, you are left with difficult predictions and that's certainly not the policy path to be.
8:28
We have a lot of use of enzyme production [inaudible] history with developed enzymatic protein production process and it comes now in favour of our mother company.
8:43
If we can together produce those antibody solution enzymes at the scale we want.
8:51
So with this summary formulas to clear that we can unlock the power of AI building on few mathematics that you find the on our internet, you can apply more chemical inspired workflows.
9:12
So we need a lot of chemical knowledge to it.
9:16
This is still something that I think that is needed for predictive powers and our biochemical background and inside production capability is also very important element next to simple in silico predictions.
9:33
And I hope if this I could give you a little bit of flavour about our background, what we are doing, why we are doing it, and that it's not just AI, it's AI with a human logic, human education behind which is the path to success for us.
9:53
You can talk to me at booth.