0:01
As mentioned, my name is Barry Duplantis.
0:04
I've been with Ailux for about a year and a half now.
0:07
Prior to that, I spent about 10 years in antibody discovery and engineering from both the technical and business side with about 8 years prior bouncing in and out of the biotech startup world.
0:17
So that's my quick little intro, but it's my pleasure to kind of introduce Ailux and also go through three different case studies that focus on AI powered immune repertoire mining and also multi objective engineering.
0:36
So before I jump into kind of the meat of the case studies, I'd kind of like to introduce Ailux a little bit.
0:41
So Ailux is the biologics division of a company called XtalPi.
0:45
XtalPi has been a global leader in AI drug discovery.
0:49
It was actually started about 10 years ago from three postdocs coming out of MIT.
0:54
They were originally focused on small molecule structure prediction before moving into small molecule discovery, solid-state chemistry and automation, and then the biologics division was brought on in 2018-2019.
1:07
It's important to note that we've been able to have quite a bit of flexibility through our R&D purposes because the company's actually raised over $1.3 billion in that 10 year time frame.
1:17
So a little bit about Ailux specifically, we're made-up of 85 team members across three different global sites.
1:23
From an experimental standpoint, we're about 2/3 wet lab, 1/3 dry lab.
1:28
And we really like to emphasise this because one of the major differentiators for us is the fact that the company was truly built from the ground up to integrate wet and dry lab techniques.
1:40
We put considerable amount of resources into producing, characterising and curating biologics data sets and that has fed into our AI algorithms.
1:51
And then kind of in conjunction with the panel discussion that was going on prior is we really do believe in the integration of these two things.
1:59
We believe AI is a tool that can be used in conjunction with wet lab.
2:04
But if we're given a problem to go from A to B and a wet lab problem solves that or a wet lab solution solves that, we believe in taking that route as well.
2:13
So a little bit more from 2018, we've run 80 discovery programmes, about 60-40 split between discovery and engineering, strong track record of success with 30 strategic partners globally.
2:25
We've had 7 licencing and tech transfer deals, 2 recently with J&J and UCB.
2:30
And then obviously indicator of success is 75% of our partners have initiated second programmes or considerably expanded their partnerships with us.
2:41
From a wet lab perspective, kind of as I was mentioning, Ailux would be able to compete as a premier end to end antibody discovery and engineering platform just on that credit alone.
2:52
But what really differentiates us is how we integrate our proprietary AI engines into those processes.
2:59
So the three pillars that we really stand on are XtalFold, XenProT and Xentient.
3:04
So XtalFold is our protein complex structure prediction algorithm.
3:09
We feel this is best in class.
3:10
We have external validation.
3:11
This has been out licenced to J&J and UCB.
3:15
Essentially, starting with a primary sequence of an antibody and an antigen or two proteins, we are able to produce a high quality protein complex structure that can then serve as a starting point foundation for a lot of our downstream engineering works.
3:29
The second one is XenProT.
3:31
I like to describe this kind of simplistically kind of as a biologics version for ChatGPT.
3:37
So the same way the ChatGPT will produce a sentence, a paragraph, an assay based off of prompts and constraints, XenProT will produce new biological sequences based off of prompts and constraints.
3:50
And we call these biological sequences that produces as fit because it's often used in our multi parametric optimization programmes.
3:57
So if we have a starting antibody sequence, we're looking to, you know, solve problem A, we're not just solving problem A, we're also taking the overall developability of that protein into account while running that programme.
4:10
The third one is Xentient.
4:11
This is our predictive suite of AI tools.
4:14
So essentially a hit generation programme can produce a sequence.
4:17
XenProT will produce panels of sequences.
4:20
We use Xentient to down select, prioritise, triage and select only a handful of clones to express and test on the back end that obviously meet the multi parametric requirements that so often these projects require.
4:34
So that's kind of the three different pillars that we stand on.
4:37
So it's nice to say, but we also believe that we should show this through data.
4:41
So we've got three different case studies today that we're going to walk through.
4:45
One's based off of our philosophy on discovery.
4:48
So this is deep mining of an immune repertoire.
4:50
The second one is based off of hit expansion.
4:52
So when we have identified lead candidates from a programme, we can actually go in and structurally amplify those lead candidates through NGS data sets.
5:02
The third one is our ability to refine a lead candidate, and this is showcased through dual specificity affinity modulation, multi parametric engineering programme.
5:12
So the first case study is based off of the HIT generation and we call this Trilux.
5:16
This is essentially our platform for identifying diverse, functional, developable and humanised hits from antibody discovery programmes.
5:26
So we call this Trilux because it takes advantage of the best of in vivo, in vitro and in silico technologies.
5:34
And then the three in one comes from the fact that we're actually integrating 3 different discovery engines within one programme.
5:40
So those discovery engines are XmartFluidics.
5:42
This is our in house single B cell screening platform for both plasma and memory cells.
5:47
We have XploreSeq.
5:48
This is our NGS based repertoire prediction algorithm to identify binders from NGS data sets.
5:54
And then XpeedPlay is our NGS enabled phage display platform.
5:59
So this would be fairly comprehensive from an antibody discovery standpoint already.
6:03
But where we feel we start to differentiate is actually after this first round of binders.
6:08
So in this case study, we immunised mice with a 22 Kilodalton soluble protein.
6:12
We ran the three in one approach that produced 130 binders from XmartFluidics, 85 binders from XploreSeq, 67 binders from the phase display technology.
6:22
So at this point we can actually take the sequence information from the binders and the non-binders.
6:27
At this point, the programme and we create an active learning module that is project specific.
6:32
We can then go back into the NGS data sets to identify incredibly rare clones that are still binders at a very high percentage.
6:41
So in this programme we went back to the NGS data set, and we added an additional 52 rare binders that led us to 334 binders total.
6:49
And since this is sequence heavy at I get go, we can ensure that we're selecting for maximum type diversity.
6:56
The second point where we feel differentiate starts come in is after this first round of binding, we can actually employ our XenProT and Xentient algorithms to do high throughput humanization while also selecting based off of developability traits.
7:13
So in this situation, the down selection and then the high throughput humanization, we produce 98 humanised hits that led to 33 functional antibodies on the back end.
7:24
So how we actually go about identifying antibodies from the NGS data sets, this is a little bit pared down obviously, but we've created a multimodal graph based predictive AI method.
7:34
That's not necessarily just these sets of data that's been put up here as an example, but this has created the pre-trained model.
7:41
Once we move through a programme and we start collecting data, that data gets then fed back in to create a project specific active module that consists of obviously the information from binding and non-binding sequences to lead to our fine-tuned AI module.
7:57
So one of the questions is why are we doing this?
7:59
Well we still feel that the immune repertoire is under explored even with the advent of obviously very high throughput single B cell screening methods.
8:09
The traditional methods are still being dominated by the expanded clonotypes and singleton clonotypes can still go unexplored in these methods.
8:18
So how does it specifically rate to this programme?
8:22
We had an NGS data set that was represented by 117 expanded clonotypes and around 1200 singleton clonotypes.
8:30
With random selection that we performed in parallel, we were able to identify these rare clones at about a 5% rate compared to 36% binding rate for the entire population of clonotypes.
8:41
With our pre-trained model this jumps to 15.4% with 54.1% for all clonotypes.
8:48
But it's after we create this project specific module that we see the massive increase to 61.2% for the singleton clonotypes as compared to 75.6% for all clonotypes.
9:00
The other question that comes at me a lot and us is, you know, what is the point of looking for the singleton clonotypes?
9:07
Is there an advantage?
9:08
Is it purely an academic exercise?
9:11
We found that in this case study and a number of other case studies that we've run, it is not just an academic exercise.
9:19
So when you look at the outputs from the different platforms compared to a clinical benchmark, phage display, XmartFluidics, XploreSeq, they all produced very good antibodies.
9:28
But when you look at the dynamic range that comes out of the singleton clonotypes that we pull, we actually get a very interesting dynamic functional range from these clones.
9:37
We've seen this in terms of function, but we've also been able to see this in terms of unique specificities of antibodies as well.
9:45
When looking for antibodies that identify single amino acid substitutions, they can obviously be very low abundant clones within these data sets as well.
9:55
The second point of differentiation, as I mentioned is our ability to perform high throughput humanization.
10:00
Depending on your discovery philosophy, we do have a number of different ways that we can perform humanization.
10:06
One of them is what we call deep humanization.
10:09
Using XtalFold as a primary tool to identify the key contact residues, we're able to essentially perform a paratope graft eliminating a number of non-human amino acids from the CDRs.
10:23
But if your discovery philosophy is more focused on eliminating late stage attrition of clones or potential engineering issues downstream, we can perform high throughput humanization on all the binders coming out of the first round of screening.
10:37
To do this, it's essentially a combination of looking for the closest structural germline and analysing humanness and trying to eliminate potential back mutations.
10:47
The wet lab procedure is actually to express 2 variants per parental antibody and in this situation we're over 90% successful with threshold staying within a 5 fold affinity of the parental and maintaining a level of humanness compared to the clinically humanised antibodies.
11:06
We can do 500 a day if not 1000 a day.
11:09
Basically depends on what's coming out of that primary data.
11:14
The other differentiation part is obviously our ability to predict the developability aspects.
11:20
This slide here is basically representing the correlation between our predicted outcome and the wet lab validation.
11:26
The gold dots are shown as predicted positive.
11:29
The blue dots are as predicted poor.
11:31
And then once again, you can see that there's a very good correlation between the gold dots and a good developability outcome in the wet lab versus the blue dots and a poor developability output.
11:41
Once again, the point of this is essentially to eliminate potential late stage attrition of lead candidates as you go through the process.
11:50
So kind of in summary, obviously Trilux is our method of producing these highly diverse panels of developable and humanised clone from a sequence and functional level.
12:01
On the right hand side, we also show it's a relatively small protein that we obtained 5 different epitope bins, 2 overlapped with potential clinical benchmarks, but we also managed to identify a novel epitope that was only accessible through those rare singleton clonotypes.
12:19
So the next case study that I want to jump into, which I actually find very interesting is our ability to perform structure guided hit expansion of epitope specific and functional antibodies.
12:29
So in this case study, we were actually looking to identify antibodies that bound a conformational epitope that was formed during the formation of a homotrimeric complex, but not to the individual monomeric subunits themselves.
12:42
To do this, we actually created an epitope specific antigen using our AI guided protein design.
12:48
Happy to talk about that later.
12:50
That's not included in this talk, but essentially we used that antigen to immunise mice.
12:54
We ran through our Trilux hit generation programme and identified 11 epitope specific and functional hits.
13:01
We then went back into the NGS data set using our structural guided approach to pull out 30 structurally convergent sequence diverse antibodies that were then validated that yielded 12 epitopes specific hits and 12 epitopes specific and functional hits on the back end of that programme.
13:20
So essentially how we do this is we take the 11 functional hits from the get go as seed antibodies.
13:27
We then extract certain geometric and structural features from those antibodies, including shape, curvature, hydropathy, electrostatics and free electrons.
13:37
We then have a customised scoring feature that then allows us to go back into an NGS data set and pull out structurally convergent and then sequence diverse clones.
13:46
So in this situation we ended up with 30 hits that had that predicted structure of interest and a diverse sequence.
13:55
So when we verified that 30 hits, obviously we came down to 12.
13:58
But how did these actually compare in terms of sequence diversity?
14:02
From the first 11 hits, we had three different distinct sequence clusters.
14:07
After we added the 12 new hits, we added four new sequence clusters.
14:12
And these aren't just amino acid mutations in the CDRs, we're actually able to identify structurally similar clones that are functional from different germ lines as well.
14:23
The third and final case study that I kind of want to walk through is multi parametric engineering case study.
14:30
This is focused on affinity modulation of a dual specificity cytokine receptor through our active learning modules.
14:36
So the task in this situation was basically threefold.
14:41
One was affinity modulation task was to increase the affinity of cytokine to the receptor by threefold.
14:49
It was to decrease the affinity of cytokine B to the receptor by tenfold.
14:54
We were also tasked with maintaining overall developability comparable that to the wild type receptor.
15:00
But then from a patent perspective, we were tasked with identifying novel mutations that weren't present in the IP space.
15:08
So to begin this programme, we were fortunate enough to actually have a crystal structure available for cytokine A and the receptor, but we did not have a crystal structure available for cytokine B and the receptor.
15:18
So we performed our XtalFold structural analysis on this, which then basically led to the identification that yes, in fact these two cytokines were sharing a dual binding interface that had a potential of 110 contact residues on the receptor.
15:32
Fortunately, after the programme, the crystal structure for the cytokine B and the receptor was released that allowed us to benchmark our prediction which yielded a DockQ score of 0.667 and an IRMSD of 1.62.
15:46
So this was considered a successful prediction by our parameters.
15:51
So essentially how we go about this multi parametric optimization is very similar to how we tackle a number of different engineering programmes.
16:00
Essentially we rely on XenProT as I mentioned to produce these fit sequences and drastically reduce the overall mutational search space.
16:07
So in this situation, one of the constraints that was placed upon XenProT was obviously the 110 contact residues, the patent space.
16:15
And then in the first round of engineering, we limited the number of potential mutations to a single amino acid per construct.
16:22
So we had XenProT produce these sequences and then we then down selected and ranked them using the Xentient predictive AI, at which point we're producing between 30 and 50 sequences that are then gone gene to protein and we test in the wet lab.
16:38
So in this situation, we did two point SPR and then also tested the yield impurity to ensure overall stability of the protein.
16:46
We then take the data that we've learnt from this round of information.
16:50
We create an active learning module specific to this programme and then we repeat the cycle again looking for improvements.
16:58
So the overall output of the programme took 4 rounds to yield 7 cytokine mutants that had the desired profiles that we were looking for.
17:06
I do find it interesting and very realistic that you can see the progression of learning throughout the different rounds that we're looking at.
17:12
So the white dots represent round one where mutations were limited to a single amino acid per construct.
17:19
The yellow dots represent round two.
17:21
Once again we limited the XenProT algorithm to making a single amino acid mutation per construct.
17:28
And then where we saw the large jumps was round 3 and round 4 where we increased the number of potential mutations that the algorithm was able to produce to 9.
17:36
So the final results from that round 3 and 4 all had between two and six point mutations per construct to hit the target profile that we were looking for.
17:48
So in summary of this, basically we leveraged all three of our engines through 4 rounds of optimization.
17:54
We expressed 233 variants and yielded 7 patentable novel variants.
18:00
So we do consider this one of the more complex or a more complex engineering programme that we've run.
18:06
We can use the same principles to run more simplified affinity maturation programmes.
18:11
This is just a very quick snapshot showing the affinity maturation of a Fab that started from 5.78 nanomolar.
18:18
We increased to 0.445 nanomolar 13 fold increase.
18:22
This was two rounds of optimization, 65 constructs, 5 weeks.
18:26
And once again, you can see the progression from round one to round two in terms of our learning and our ability to produce sequences that more match what we were looking for.
18:35
So this is just three different case studies that we have.
18:37
We have a lot of different case studies focusing on a lot of different aspects that we're happy to walk through after the fact that the booth if you'd like to come by.
18:45
But otherwise, I'm happy to take some questions now.
