Executive Interview with Vladimir Chorošajev & Cédric Bény, Cortex Discovery

0:03
Hi everyone and thank you for joining us today.

0:06
I will be interviewing Vladimir, a machine learning researcher and Cédric Bény, a Co founder and CTO at Cortex Discovery, a company developing deep learning models to predict therapeutic effects of target effects and pharmacological properties.

0:22
So Vladimir and Cedric, thank you both for joining us and taking part.

0:26
So I guess first question is, could you briefly describe what problem Cortex Discovery is addressing in drug discovery?

0:35
So we are trying to address the problem of drug discovery.

0:41
We want to change quite a bit in the way that the drug discovery process is done in mostly in the preclinical early stage of drug discovery.

0:50
Especially we want everything to happen in silico first on computers.

0:55
We want it to be automated.

0:58
We want the ability to systematically explore the whole space of molecules.

1:03
And we want the drug candidates that come out of this process to have a very high chance of success once they are synthesised and tested in the lab.

1:13
And to do this, we are building a large neural net, a foundation metal for drug discovery, which we train on all the hard data that we can find, which for now includes any lab measurements of biological activity of small molecules in all therapeutical area, physical, chemical admit properties.

1:33
And we also developing the ability to learn from the laws of physics and chemistry themselves, which you know, which explained the way that things that the activity happens at the molecular level.

1:47
And the result from the system is that we get this large neural Nets, a single net that can make all the prediction.

1:55
Let's call it an AI which is not memorising or interpolating the data, but it is really able to discover new fundamental principles that explains the data and that makes it able to generalise to the new molecules.

2:11
Also, the neural net predicts everything at the same time, which means that we can screen molecules based on many criterion at the same time, including admet of target.

2:22
It's it still is fast.

2:24
So that means that we can quickly screen, for instance, all participle molecules in matters of minutes or we can use the system to explore the whole space of molecule to progressively build, explore new structure guided by the neural network predictions.

2:43
Also it's system is accurate.

2:45
That's very important.

2:46
That's something really a push because it makes all the difference.

2:51
So it's we, whenever we were able to test that we found it to be as accurate as experiment and sometimes even more accurate because it's able, the system is able to understand the noise in the data.

3:05
So a big advantage of proceeding this way is that for data reuse.

3:11
So when the companies, you know, invest a lot to make a height of the screen, for instance, you know, they will just use it to find some molecule and mostly discard the data afterward.

3:23
But the system is data can be integrated into the AI not only to make it able to predict new molecule for the target, but every day that we add also improve the model predictions for many targets because it improves general understanding of biochemistry.

3:40
Perfect.

3:41
Thank you very much.

3:41
And when I was talking to Vladimir last time, he mentioned that Cortex Discovery has achieved a couple of major breakthroughs recently in the antiviral and oncology space.

3:53
So could either of you maybe share like one or two examples that highlight the power of your models?

4:01
Yeah.

4:01
So in the field of oncology and virus, this was mostly for the first projects we did to as a proof of concept to test that the system actually works in the lab.

4:12
But we got much, we actually got useful results from this.

4:16
So the first thing we did was it was still during the COVID, late COVID time.

4:21
We, it was the first time we actually tested the medical in the lab.

4:26
It was a collaboration with Aarhus University in Denmark.

4:30
And the researcher that we worked with was able only to test a few molecules, 6 molecules.

4:37
So we send them our six best candidates for repurposing candidates.

4:43
So drug that already exists, but to see if that can actually be useful for a different reason than what they were sold for, namely in that case for inhibiting the SARS-Cov 2 virus.

5:02
And so out of the 6th candidate, we sent our collaborator five of them that are to be active at the predicted concentrations and two of them were highly potent.

5:13
And we pursued these two compounds, they were shown to be active in other, in many different cell types and also in 3D human cell models.

5:25
And we've a patent for this and we also have a published academic paper in the journal Antiviral Research.

5:34
Another project we did also on viruses that the second project we did, it was a collaboration with SINTEF and SCMM in Norway.

5:45
Here we decided to pre commit.

5:47
So we just want to target GP 41, which helps the virus enter the cell, the HIV virus.

5:54
And we wanted to have a way to really prove to other people that, you know, we are doing real predictions and not both predictions.

6:04
So we pre committed the top 10%.

6:07
So we screen virtually library, a small library of 2000 molecules.

6:11
We pre committed our top 10% candidate ranked by probability of being active to them before the experiment.

6:21
And then they did the experiment and they found that they found 22 hits of which 21 were in our pre committed list of top hits.

6:32
And the we were able to use this to make an experimental AUC.

6:36
So a measurement of how accurate our prediction are.

6:39
We obtained 95% which is extremely good and which was close to what we predicted using our own testing.

6:48
We also did a quick project with the VCFG Centre Oncology Research Centre in Australia, and here they shared with us some data of toxicity for cancer cells, specific cancer cells, and they kept a number of unpublished results for themselves to be able to check that our predictions were correct.

7:13
And that way we also were able to evaluate our accuracy.

7:16
And we found that we were on that one as accurate as the experimental accuracy in the sense that they had two different data sets on the same target.

7:26
And we were able to use that to measure how repeatable their experiment is.

7:30
And the AUCB that give us a sort of experimental AUC accuracy.

7:35
And ours was equivalent to the experimental 1.

7:41
So these are our first 3, three main projects that convinced us that our system is really working.

7:47
Very exciting indeed.

7:48
And you mentioned that you're doing a lot of collaborations at the moment.

7:52
So I guess what kind of companies and research groups are the best fit for you to partner with at Cortex Discovery right now?

8:02
So right now, pretty much any early stage discovery team, whether it's CRO or a research group or a pharma company, we can have them at any stage because we can benefit from the thousands of predictions we can already make with our data.

8:20
If they have data, we can train on it as well, integrate it into the system.

8:25
And this you know, we can train our system pretty fast.

8:27
So we can train it in a week, which means that we can create model specific for customers without having to share data across customers.

8:37
And yeah, and the even we don't have data for us.

8:41
The pre training system can also be used for if you don't have the target that they are interested in.

8:46
In our data we can use the system for off target and admit labelling of compounds.

8:54
Also we have a way to if a company does high throughput screens, we have a way to if you combine any HTS with our system we can really supercharge the HTS.

9:06
We made some estimates that we can deliver 20X more value, meaning we can for each hit we can obtain each hit at 20 times less cost and also increase speed.

9:22
We also look for any partners that will help us develop our longevity pipeline.

9:27
So we do not on our side, we only do in silicone.

9:30
So we need partners to also test the molecule in the level.

9:35
That's about it.

9:38
Perfect.

9:38
Yeah.

9:38
And I guess that was coming on to my next question that you guys are also exploring the longevity and the longevity space and DNA repair targets and what are the opportunities that you envision in the longevity space for you and your clients?

9:57
Yeah, so in the space of longevity, we are currently focusing on the mechanisms of DNA repair.

10:05
So whole molecular pathways and whole systems, for example, double strand break repair, mismatch repair, etcetera.

10:15
So we think that the DNA damage is the primary cause of other downstream effects that happen to have very bad effects on our general health.

10:29
So it's the core driver of other things that are tackled to the longevity space, which are like senescence, inflammation, etcetera.

10:38
So we try to kind of tackle the problem and the root here and by focusing on the DNA repair systems.

10:45
So we currently have a partnership with Vienna University and NIH where we are testing 4 candidate compounds and one of them in animal models already, so the mice.

11:00
So besides the purpose of the longevity research being valuable in itself, we also think that it's very promising as an agent to cancer therapy because a lot of oncological compounds are damaging DNA, which leads to secondary mutations to relapse to genomic toxicity.

11:28
And we think that this is something very that's very important to prevent by supercharging our own DNA repair mechanisms.

11:37
The second very important goal here is that the longevity research is quite a dark forest in the whole field of pharmaceutical research.

11:47
That means it's not very well known, but should work.

11:53
It's there's not a lot of data, not a lot of publications.

11:57
And we generally cannot buy DNA repair drugs in your local pharmacy, right.

12:05
So we also want this successes here to be a strategic demonstration of an AI system tackling biological problem that's very has very scarce data and very scarce intuition of how things are supposed to work.

12:22
We want to really show that the machine learning models here can discover something novel we never thought about in a very unresearched space.

12:33
I see.

12:33
And do your clients often, what are the kind of main misconceptions they have about longevity, or do you face a lot of scepticism when you present these potential solutions to them?

12:49
So we try to force their scepticism.

12:54
Longevity, as it feels, has an interesting history, but we try to tackle scepticism with basically hard data.

13:03
So for example, like one of our very recent achievements is an AI generated compound that increased the lifespan of a whole Organism, in this case worms, C elegans worms by 30% medium.

13:20
That's not a lot in human time.

13:23
That's basically going from 16 days to 21 days, but that's still a very massive improvement for that was driven entirely by machine learning pipeline without much human expert knowledge in the loop.

13:40
So we, I think that the way to fight scepticism in these domains is just by presenting very hard experimental evidence, preferably life models.

13:54
Perfect, thank you very much.

13:55
And do you find that large pharma companies and biotechs are receptive to integrating AI models into their existing discovery workflows?

14:06
Are there any kind of organisational barriers that you come across?

14:10
So this is a very difficult question because on the top down level, pretty much any big pharma company nowadays is claiming that AI is a strategic priority for them.

14:25
But we think that this change doesn't really come from the bottom up because the actual researchers are still very often using old school work flows and doing things the way they did.

14:41
And we are really trying to focus this problem by focus on this problem by making useful tools that are useful for the researchers bottom up.

14:51
We want that people working in the actual drug discovery space make the decisions to use our tools and to push them to their management and not the other way around because the top downs initiatives are often perceived as artificial in any kind of company.

15:08
That applies to software, that applies to pharma, to everything I see.

15:14
And in terms of ensuring model generalizability, a special especially with sparse experimental data in novel chemical spaces.

15:25
How do you ensure this?

15:27
So our approach is the foundation model approach.

15:32
This means that any of our models are never trained on a single task, which means they do not have the capability to memorise the data and to come up with some curve that feeds the data but does not reflect the physical reality.

15:53
So we architectural, we make the model architectures in a way that the models are forced to achieve understanding by compressing information.

16:05
So instead of memorisation, they are forced to derive the actual biological principles and chemical principles that the molecules are working on.

16:16
So in this case that data scarcity isn't that big of a problem because models make their decision based on the laws of nature, the physics, the chemistry, the biology of the interactions and not by looking at the supply data that much.

16:37
It just needs a little bit of data to understand which physical principle to apply to the specific molecule.

16:44
We have the prior knowledge that fills the gap, involves understanding and don't need large data sets for small tasks.

16:52
Perfect.

16:53
Thank you.

16:54
And we've touched a little bit on the importance of off target data.

17:00
Could you maybe expand a bit on how Cortex Discovery’s ability to obtain this off target data benefits clients, particularly in the early phases of drug discovery?

17:13
So basically in the drug discovery process, whenever you get a number of drug candidates, the drug candidates are filtered out in the later stage based on different properties of the molecule.

17:33
So you don't only want your molecule to be active for some specific target or pathway, it's also very important that your molecule isn't active where it shouldn't be.

17:44
This is a driver of toxicity.

17:45
This is a driver of unwanted interactions and side effects.

17:50
And in the normal pharmaceutical workflow of target effects require separate experiments.

17:58
And we think that we can really cut some time here and some expenses because our models as foundation models, they predict everything at the same time.

18:09
So whenever we're trying to find the activity for some target, we also get all the other activities, every single one of them at the same time.

18:21
That's great.

18:22
Thank you.

18:23
And in the next one to two years, what are the main goals for your company?

18:30
Yeah, so at the commercial level, we want to scale adoption.

18:33
So get more customer, get our system, get them used.

18:38
We also developing our longevity pipeline and on a tech label we are developing also.

18:50
So right now we, our system, we use ourselves and we interface our system by providing a service, but we also working on creating user facing interfaces that people can directly access our own interfaces, which will allow us to scale and get more customers.

19:09
And scientifically, something very important that we are doing now is to develop this, the ability of the model to use physics and chemistry laws instead of just data.

19:23
Essentially it's based on the physics based neural binding free energy model.

19:28
So we use neural Nets to be able to predict from knowledge of the physical of the chemistry, to predict the affinity, the dissociation constant of small molecules and protein target.

19:43
And you want to do this right now, you know, this kind of, this is done by simulations, but these simulations are very slow.

19:50
So we are we have assumed that we make this millions and times millions of times faster, which will allow us to also integrate the system with our main neural net and to screen also.

20:02
Any targets that for which we don't even have that data or even or small amount of data that could really greatly expand our capabilities?

20:12
Very exciting.

20:13
And how do you picture the, I guess on a more broader level, the competitive landscape evolving for AI drug discovery start-ups over the next, let's say over the next few years?

20:26
So that's a very interesting question because currently we see that the field of AI for pharma is very much flooded with language model approaches, so basically putting in some wrapper or fine tuning on any kind of large language model.

20:47
And we think that this is a bit of a dead end here, because our core belief here is that molecular biology isn't structured like a language.

20:58
And any approach that will truly see success in giving accurate predictions need to reflect the structure of the physical world in the architecture of the models.

21:10
So working with 3D molecular structures, working with potential forces, physical constraints.

21:18
So we want the models to get the physics of binding, physics of molecular activity, not from text written in articles by very smart people, but from actual equations.

21:36
So we see that physics inspired architectures, for example alpha fault or bolts have already achieved quite some leaps in what they can do their breakthrough, but they are still dependent on data availability.

21:53
And we think that the next breakthrough in the competitive landscape will come when we are able to train the models not on data, but on the underlying equations and physical principles directly.

22:07
So this is an area of our focus right now with the physical model.

22:12
And we also think that whatever company achieves learning from the underlying principles directly can dominate the competitive landscape.

22:25
Very interesting insights.

22:27
And I think I have one final question probably for you, Cédric, I guess on a personal level and what has been the most rewarding aspect for you as a Co founder and CTO, I guess leading this company through probably some quite drastic transformations?

22:46
Well, I mean, the ability to create a system that actually has an impact in the world and to see the predictions work in the lab.

22:55
I mean, the greatest moment was the first experiment we did because we had everything in silico and the computer and we had, we thought that the same work, but we had no idea that whether it would actually work.

23:06
Once we face reality is hard.

23:10
And so the fact that it works so well was really amazing.

23:15
Also, it's a big change from my past as I used to be a researcher in theoretical physics.

23:20
So everything I was doing had absolutely no relation to, I mean, it was trying to model the real world, but there was no actual observable consequence of my discovery.

23:31
So this is a big change.

23:34
Very good answer.

23:35
Well, I think that's everything from me.

23:37
So Vladimir and Cedric, thank you very much for sharing your insights and a bit about your work.

23:42
Cortex Discovery definitely sounds like an exciting time for you guys and the company and for the viewers.

23:50
Thank you for listening and please do stay tuned for more conversations.

23:55
And Vladimir and Cedric, thank you again.

23:56
Once more, thank you very much.