0:00
Currently serves as the head of AI Drug Discovery at XtalPi.
0:05
In this role he's pure heads a dedicated team in the development and application of innovative AI and physics based methodologies for advancing drug discovery projects.
0:15
His academic background is in computational chemistry, and he holds a PhD from ETH Zurich.
0:21
So he will tell us all about accelerating drug discovery with AI and next generation automation, so here we go.
0:28
Thank you for the intro and thank you all for coming to the talk.
0:36
It is my great honour today to be presenting here on the behalf of XtalPi team.
0:44
I'm going to introduce a little bit about our platform using accelerating drug discovery, using AI and automation.
0:57
So I'm going to talk a little bit, introduce a little bit about our company XtalPi.
1:02
And then majority of my time will be spending on our platform of AI drug discovery, and I will touch a little bit on the automated chemistry and biology platform.
1:16
So about our company, XtalPi is a company that is aiming to develop the new paradigm for future drug discovery.
1:26
We are a science and research driven company.
1:29
We have four sites globally and we have around 700 people with 70% of them belonging to the science and tech team and we have a large lab space as well.
1:45
So we've built various different platforms including the small molecule drug discovery platform as well as the solid-state platform which is used for crystal structure prediction of small molecules.
2:01
And we have automation platform as well as the antibody platform as well.
2:07
But today we will be focusing on the small molecule drug discovery.
2:13
And here are some examples of our recent collaborators and we have strong track record in partner with global pharma.
2:22
And in 2018, we started our cooperation with Pfizer for solid-state prediction.
2:29
And in 2022 when they were trying to push Paxlovid, their COVID-19 drugs, we did the crystal structure prediction for that particular drug and it took us about six weeks to finish the predictions and that kind of accelerate the process of that drug coming to the market.
2:53
And in 2023, last year, we initiated a partnership with Lilly, and we got very good feedbacks from our client.
3:02
And I'm not going to read all that.
3:06
So for the next part, I'm going to talk a little bit about our drug discovery platform.
3:15
And as mentioned, the vision of XtalPi is always to build a new paradigm for future drug discovery.
3:23
And we try to achieve the goal of design smarter using AI algorithm.
3:33
And we try to achieve the goal of make faster using automation.
3:39
And our goal is to of course, to success the to improve the success rate of drug discovery process and of course to achieve the ultimate goal of bringing effective medicine to our patient faster.
3:57
And we try to deploy AI and automation platform where they actually access the most.
4:03
So we provide solutions to drug discovery problems using AI algorithms and we also provide actual library synthesis using AI and automation.
4:19
So this slide shows some of our core values.
4:23
So first of all, it is superior accuracy and diversity.
4:28
We were able to achieve 90% of success rate for our hit and lead discovery in over 60 projects that we had in house.
4:38
And the second one is test fast and succeed fast using automation platform.
4:47
And with those platforms, our goal is to tackle challenging target.
4:56
So for the next part, I will introduce a little bit of our AI drug discovery platform.
5:04
And I'm going to start with a case.
5:08
This is a recent collaboration that we did with one of our partners.
5:13
So our partner had a target and this target, it binds to the allosteric site and it is very flexible and that kind of introduced some specific challenges for computational modelling.
5:27
And on top of that, this is a very hot topic, very hot target, and a lot of companies are going after that and they are a lot of patents published from various different companies.
5:40
And why I was checking all those patents, I had the feeling that didn't I just look at that exactly the same patent before.
5:46
So some of the human designs based on our experience might fall in the scope of other patents already.
5:54
So that's when we started to work with our clients and we, this middle part is the approach that we actually take.
6:07
This is the approach that we take.
6:08
We generate compound, a million compound using AI generated models.
6:15
And this was just the first step, right?
6:18
We still need to select the right compound from these million compounds.
6:24
So we did some of physiochemical property filter and that still we have 100,000 compound left, and of course we dock all these 100,000 compounds.
6:34
And this is when we utilise our XFEP platform to assess a very diverse set of compounds.
6:45
So we were able to assess 5000 compounds in a single month.
6:49
And imagine that if we were using some other commercial software which is limited by the number of licence, the throughput would be a lot lower.
6:59
So maybe we can assess about 100 compounds here.
7:03
But then the diversity of the compound that we were able to assess would be a lot less.
7:09
So we were able to assess 5000 compounds in a single month.
7:13
And then according to the FEP score, we were able to select around 100 compounds.
7:19
Then we sit down with our partners and then we select together about 30 compounds for synthesis.
7:25
And in the end, twenty of those compounds were made and 40% of those compound were actually having potency better than 100 nanomolar, which were classified as high potency according to our partners.
7:40
And what's more important is that there are 10 different scaffolds within these 20 compounds and we did all those calculations in just a single month followed by the following synthesis and testing of course.
8:01
So here are a little bit more of the results that we got.
8:04
The left hand side shows the diversity of the compounds.
8:09
The X and Y axis are the projected chemical space and the further the distance on the plot, we can consider that those compounds are actually more dissimilar.
8:23
So we can see those orange dot in the middle that were the original set of compounds in our client they sent to us.
8:30
Those were actually very similar to sound of a compound in the in other patterns as well.
8:37
And the part that we were trying to modify is the pyridine part.
8:40
So as we can see that we were able to get various different scaffold including some saturated ring, some fused rings.
8:48
So the diversity of compound is a lot larger than the initial set of the compound which of course outside the scope of patent that are covered already.
9:04
On the right hand side, and those are all selected by as I mentioned the FEP calculations.
9:09
And the right hand side shows the comparison of the FEP performance for this particular target.
9:14
This is a head to head comparison between our software and the commercial software, and for exactly the same set of compounds.
9:21
And of course in this particular case, we our performance is slightly better than the commercial software.
9:29
So with this case study, what I wanted to show is that the strength of our AI platform is the accurate and efficient exploration of very diverse chemical space.
9:50
And this will all done by the combination of AI and physics space modelling.
9:57
Some companies claim that they can do drug discovery using AI and we took a slightly different approach here because we understand that artificial intelligence is very powerful in terms of drug discovery, but it also has its limitations, right?
10:15
The AI models may be accurate if our predictions are kind of similar to the training set and it requires a huge set of good training set.
10:29
And AI models, it is kind of creative, but it's in the sense that it's kind of a black box model.
10:35
And in that sense, actually physics based modelling can be a very good complementary to those AI methods because physics based modelling is first principle and does not require any training set.
10:51
It is of course lower than AI algorithms, but it is more accurate in terms of, for example, binding affinity prediction.
10:59
And in addition it can provide a structure inside which is complementary to kind of black box predictions of AI algorithms.
11:13
This is an overview of our AI platform that were used in the case study that I just showed before.
11:23
So the left hand side shows the AI generated models.
11:26
Nowadays there are so many different AI generated models.
11:30
So we build our platform such that it is customised for various different drug discovery scenarios like first in class, best in class, and as mentioned that AI can easily generate a million compounds and what is as important as generating those compound is to the select right ones out of those million compounds.
11:52
So we utilise on the right hand side active learning, try to speed up this process and according to our own test and according to the literature, it is able to speed up the screening process by at least one order of magnitude.
12:08
And as other companies we also build our in house [unclear] property prediction model to facilitate the screening process.
12:20
This is the summary of our physics based modelling platform.
12:25
And as some of you may know, the Force Field is the fundamental factor of all the classical mechanic simulations.
12:36
And in terms of that, we actually have built a state of art Force Field because we started our collaboration with Pfizer in 2018 for the Force Field and now we have a very good Force Field and that's leading to the accurate binding affinity prediction.
12:52
And as we know, a pose is also very important for all the structure based drug discovery process.
13:00
So we've built an XPose platform which utilises extensive molecular dynamic simulations and enhanced sampling techniques to find a binding pose as accurate as possible. And of course the binding affinity prediction: we've built an XFEP platform, and it has several different features.
13:22
The first one is quite comprehensive in terms of functionalities.
13:30
So apart from the traditional relative binding free energy that people usually run with FEP calculations, we have implemented absolute binding free energies, scaffold hopping, protein mutations, as well as for example covalent inhibitors. And it is able to handle various different kind of drug discovery scenarios.
13:55
It is also accurate.
13:57
So we did a head to head comparison between XFEP and the commercial software.
14:05
And according to our test on this dataset, our performance is at least as good as commercial software if not better.
14:14
And another feature of XFEP is that the throughput is a lot higher because we are not limited to the number of licence that we have to run FEP calculations.
14:26
So we can get through a lot more compound with the XFEP platform.
14:33
So to summarise everything up, our physics based platform for binding affinity prediction, it is accurate, it is efficient and high throughput, and it is comprehensive in terms of functionalities.
14:48
And last also very importantly, it is validated in a lot of our public data set as well as our internal pipelines.
15:01
So now I'm going to give you a quick case to demonstrate the accuracy of our physics space platform.
15:09
So we work with a partner where they are actually working on very popular targets and they wanted to do some blind test of binding affinity of the target of the size of the target that the data that they have in house.
15:32
So they find some different vendors, they send all those vendors including us the compounds and they ask us to do a blind test.
15:42
So what we did was that we finished the blind test within 10 days.
15:49
And the first step is that we try to look for in patent some compound that is very similar to the chemical series that the partners shared with us.
15:58
And then we did a retrospective analysis.
16:01
That's the plot that we showed here.
16:03
The X axis is the actual data that the pattern has.
16:08
the Y axis is our predictions.
16:09
As we can see that the correlation between the prediction and data is very good. And that gave us some confidence on the prediction of the compounds that our partners actually provided to us.
16:25
So we did the predictions and then we gave the data back to the partners and this is the result that we got in the end.
16:33
Again, the X axis is the data now it does not have an absolute value, but it's in range and Y axis is the predicted 1.
16:44
And as we can see that we can certainly see a very good trend between the actual prediction and the data and we outperform all the other vendors.
16:55
And of course, in the end we started this cooperation with our partner for this particular target.
17:07
I think that's all I would like to tell you.
17:09
And I'm going to skip this part because I still want to touch a little bit on the automation part.
17:17
So to summarise this part up is that we try to combine the strength of AI method as well as the physics based modelling to achieve the goal of accurate and efficient exploration of chemical space.
17:36
So now since I still have a couple of minutes left, so I would like to touch a little bit on the automation part.
17:44
So our automation actually started in 2020 in our CDO's home, and it was built using Lego and 3D printing.
17:53
And in 2021, we already have our first automation island, which includes 5 of the prototype workstations and one IGV.
18:02
And now in 2023, we already have 200 workstations in various locations.
18:10
And this is how our automation lab looks like.
18:13
We have on the left hand side is our workstation area and on the right hand side it is our scheduling and controlling area.
18:22
And our ultimate goal is of course to build an end to end automation platform, but we also emphasise the human machine interactions.
18:35
We built not only the workstation but also software.
18:41
And on top of that we are developing AI algorithms as well.
18:47
And of course we emphasise the importance of data and IP protection.
18:53
And we utilise all that cool machine and software and AI algorithms to do automatic synthesis, which is of course more relevant for drug discovery process.
19:05
And it is able to cover 80% of the common medicinal chemistry toolbox.
19:10
And that will help us to achieve to do rapid library synthesis with superior novelty and diversity.
19:21
And we provide focus library for specific target as well as fixed library for particular therapeutic area.
19:31
And for the sake of time, I'm going to skip the case studies for automation part.
19:37
And just to mention that we do partnership.
19:41
We specialise in chemistry and if you are strong in biology, I'm sure we can find a lot of good synergy between us.
