0:00
I'm an alliance manager at Ailux Biologics and the title of my talk today is ‘Cutting Through the Hype: Four Practical Cases of AI Empowering Antibody Discovery’.
So,13
So everybody's talking about AI these days.
0:15
There's lots of talks at this very conference about this topic.
0:19
And so people are making a lot of different claims about what AI can do for us in antibody discovery.
0:25
And so at Ailux, we wanted to ask the question, is AI really having an impact or is it a lot of hype?
0:33
And so in order to address this question, we wanted to qualify it with some rigour.
0:39
And so we wanted to look at some case studies and assess the relevance of using AI in antibody discovery.
0:49
So how often do we want to run into the problem of replaceability?
0:53
Is AI the only approach that we can use are there other alternatives?
0:58
And then finally, in terms of solutions, can we successfully repeat the results we see using AI and antibody discovery?
1:08
So today, I'm not trying to give you any hype.
1:10
I just want to present to you for case studies of how we have incorporated AI into our antibody discovery workflows at Ailux.
1:18
And so the first case I'm going to be sharing with you is related to antigen design.
1:22
And then we'll talk about a case of hit generation.
1:26
And then I'll share with you two cases around lead optimization and engineering.
1:32
So the first case is a situation where we had a very challenging GPCR target.
1:38
So it's a class A GPCR.
1:40
As you can see here, it's got a very small extracellular domain, so there's little real estate for antibodies to bind to and it has low homology with cyno.
1:49
There is an existing clinical benchmark antibody, but it was only obtained after screening 300,000 hybridoma and it's not cyno-cross-reactive.
1:59
And so our goal was to generate a new panel of antibodies to this GPCR immunising with multiple antigens, so protein, cells, and DNA, to increase our chances of success.
2:11
However, the wild type protein not surprisingly does not express well.
2:16
And so here we were able to apply our XrossXeven platform which is an AI guided GPCR antigen design with large scale mutation.
2:26
So this slide is showing you a schematic of the general workflow.
2:30
So we start with the sequence of the wild type GPCR, and we subject that to our generative AI model that creates a large panel of mutants.
2:41
And I should note we're not mutating the extracellular loop, just the transmembrane and intracellular regions.
2:47
From there we can use predictive AI and molecular dynamic modelling to pare down the panel that we think will be the best representatives of the wild type protein and then express a small number in the wet lab and validate them.
3:03
So at the end, we end up with AI design GPCRs with anywhere from 120 to 160 point mutations, so quite large scale mutations again, only in the TM and the intracellular region.
3:19
So to dig a little deeper into the AI models that we're using with XrossXeven, the first being a generative AI model called XenProT.
3:28
And here what we're able to do is generate mutants that are predicted to fold into the wild type confirmation and behave like a natural protein.
3:38
From there, we can use predictive AI ranking using two of our algorithms, XtalFold and XcelDev.
3:45
And I'll talk more about these two later on, but we're able to rank them in terms of the extracellular confirmation in comparison to the wild type and also look at biophysical properties in silico.
3:59
And then finally, we can use MD modelling and rank them even further and pare it down to just 15 mutants that we expressed in the wet lab.
4:11
So with this XrossXeven platform, we were able to generate GPCR mutants that express much better than the wild type.
4:19
As you can see on the left, that's an SDS-PAGE with two mutants, NS4 and NS6.
4:25
Compared to the wild type, they were thermo-stable.
4:28
And then we also tested them for binding to the benchmark antibody.
4:31
So they are properly folded.
4:36
But the real test is immunising with this protein antigen.
4:41
So we took this mutant GPCR immunised mice and we're able to generate a strong titer as shown on the left.
4:47
This is binding on cells expressing the wild type GPCR protein.
4:52
And then we did a very small scale screening and identified panel of 13 hits that again binds cells expressing the wild type GPCR and that included one antibody that was cyno-cross-reactive and had comparable function to the benchmark antibody in an in vitro assay.
5:09
So if you recall the benchmark antibody, it took 300,000 hybridoma to find that one clone that was not cyno-cross-reactive.
5:16
We had a panel of just thirteen hits and we're able to identify a cyno-cross-reactive and functional antibody.
5:24
So to recap and to apply those parameters I talked about at the beginning in terms of relevance, I think using AI here is very relevant because multi-pass membrane proteins are important targets including GPCRs and ion channels.
5:39
And we all know that they're notoriously difficult to produce.
5:50
In terms of replaceability in AI versus more traditional methods, I don't think AI is replacing other methods here.
5:54
And indeed we did immunise mice using other methods including whole cell and DNA.
6:02
So I think using the AI designed mutant protein here is a complement to the other approaches that we're taking.
6:09
And in this case, the mutant GPCR protein gave us the best serum response.
6:17
And in terms of generating GPCR mutants, of course, people have been doing this for a while, but it relies on trial and error and it's very time consuming to make all these different variants.
6:27
And our XrossXeven platform can generate mutant proteins in eight weeks, whereas in some cases doing it the more traditional way and testing empirically could take a year.
6:39
So it definitely speeds up the timeline.
6:42
And then in in terms of reproducibility, we've able been able to apply the XrossXeven platform to other targets including CCR8, which is also a class A GPCR.
6:53
So moving on to our second case study, which involves hit generation.
6:59
The challenge with this target was that in previous campaigns, we had a low functional hit rate, and we needed different levels of functional activity.
7:08
So in comparison to the existing benchmark antibody, that antibody demonstrates excessive functional activity, which leads to toxicity.
7:18
And it also has a suboptimal developability profile.
7:22
So our goal with this project was to generate a new panel of hits demonstrating a range of activity levels and hopefully less potent than the existing benchmark.
7:32
And of course we wanted a favourable developability profile and we wanted to complete this in a short time.
7:39
So here we coupled NGS sequence analysis with AI in our XploreSeq platform.
7:48
And so with XploreSeq, the basic workflow is that we harvest material from an immunised source and then we perform NGS analysis both in bulk and at the single cell level.
8:01
We then feed this sequence data into our AI engine, which allows us to perform hit prediction, and then we can narrow it down to a small panel of antibodies that we can express and test in the lab.
8:14
And nice thing about this workflow is that we can use the data that we generate from this first round of XploreSeq analysis, feed that back into the AI model, and then identify additional hits in a second round of analysis.
8:29
So the whole process can take a minimum of six weeks.
8:32
So it's quite fast and we're able to identify diverse panels of hits because we're looking at the sequence data upfront. And in parallel with hit prediction, we can also select for favourable developability, which I'll talk about in a minute.
8:47
And overall, this allows us to achieve a high hit rate.
8:52
So the question becomes, when you're dealing with large sequence data sets from NGS analysis, how do you predict hits?
9:00
I mean, if you look at the sequence space from a typical immunisation campaign, you're looking at, you know, millions of unique VH and VL sequences.
9:09
And so this is where AI can come in.
9:15
But in terms of selection models, there are two fundamentally different types that we can employ.
9:21
So here I'm using an analogy to facial recognition software and with a heuristic model.
9:28
You can think of this type of selection model as a hand-picked feature.
9:34
So things like eyes, nose, mouth, things are that are easily recognisable and understandable to the human eye.
9:41
However, if we use an AI model, there are lots of other features that are used in this selection model, and a lot of them are really beyond human understanding.
9:54
But what we can do is step back, let the computer do the selection, and that's why facial recognition software works so well.
10:03
So if we apply this same thought process to antibody discovery and selection, we've been using the heuristic model for quite some time now.
10:14
And so if you think about hand-picked features, these would be things well known in immunology.
10:20
So for example, frequency of a particular sequence, somatic hypermutations or isotype looking for class switched sequences.
10:31
And based on this type of model, you would select sequences like #2 or #6 where you see an over representation in the population.
10:41
And so this model does work, and you are able to get hits that way.
10:45
But we're proposing that AI could provide a better way to do this.
10:49
And so, what we've developed at Ailux is an AI model where we've taken large amounts of NGS data and looked at not just the hand-picked features that I just showed you, but a whole array of over a thousand candidate features and used these for hit selection.
11:11
And what we've been able to do is, through empirical optimization, refine the model resulting in a model where we use over 100 features for hit prediction and selection.
11:24
And so we have done head to head comparisons of the AI model with heuristic selection model on the same set of NGS data.
11:35
So on this particular data set, using the heuristic model, we had a hit rate of 10%, whereas with the AI model, the hit rate was over 50%.
11:45
So again, this is the same set of data, but we're improving the hit rate over fivefold just by using our AI selection strategy.
11:55
And another really attractive feature of this XploreSeq approach is that we can also look at developability in silico as we're doing hit prediction.
12:08
And so we have a dedicated platform for that, which is called XcelDev.
12:12
And within the XcelDev platform, we have dedicated algorithms for all of the different biophysical characteristics that I'm showing you on this slide.
12:22
And so we can take that into account as we're sequences from our panel.
12:29
So getting back to the case study that I was talking about earlier.
12:33
So in this particular XploreSeq campaign, we were able to apply our AI selection, and we ended up with 26 clones total that we expressed.
12:44
Of those 26, 18 were binders, and we identified 4 functional hits.
12:49
You can see that the diversity of the clones that we selected is pretty high based on the germline gene usage and also looking at the CDR3 clonotypes.
13:02
And this whole process only took six weeks.
13:08
As I mentioned earlier, we can perform a second round of XploreSeq analysis based on the results of the first round.
13:14
So we had four functional hits, and we wanted to further explore that sequence space.
13:19
And so we fed that data back into the AI engine and we're able to identify 19 additional variants of those 4 original clones.
13:26
We expressed only 10 of them, but all 10 were functional hits and we were able to observe a 20 fold range of activity in our functional assay of interest.
13:40
On the left here in blue are the original 4 clones and then on the right and green are the EC50s of our round two clones in this functional assay.
13:50
And this whole process only took three weeks.
13:52
So it's a very rapid turnaround time.
13:56
And as I mentioned, we can also do a preliminary developability assessment.
14:00
So we wanted to validate those results in the wet lab.
14:03
And so we performed an array of biophysical characterization of the hits from round one and round two and compared them to our benchmark antibody.
14:15
So the benchmark is an orange and then you can see the round one hits in blue and the round two hits in green.
14:21
And overall, the developability profile of the XploreSeq hits was much better than the benchmark antibody.
14:28
And even in round two, you see some improvement compared to round one.
14:33
So this validated our XcelDev predictions in silico.
14:41
So again, to recap this particular case study in terms of relevance, I think it's always desirable to have a diverse panel of hits with a good developability profile in a short time frame.
14:52
That's what we're always striving for no matter what method we're using.
14:56
So I think it's very relevant to apply AI here.
15:01
In terms of replaceability, I'm sure there are other methods of hit selection, but those traditional methods may not offer the fine-tunability of functionally validated sequences that we can obtain with AI methods, and we can get that quick developability rating in parallel to our hit selection.
15:20
In terms of reproducibility, we've applied the XploreSeq platform to quite a few targets now, both soluble and multi-pass membrane targets, and we're consistently seeing a high hit rate in these campaigns.
15:34
So it's working well.
15:39
OK, so now I'll move to case 3, which is looking at affinity maturation.
15:45
So with this particular project, we had a parental VHH antibody with a rather low affinity of 153 nanomolar.
15:54
We didn't have any structural info on this antibody, but our goal was to improve the KD down to single digit nanomolar and again within a short time frame.
16:06
So let me just walk you through our AI-guided affinity maturation workflow.
16:15
So on the top here in the purplish light blue is our steps that are done in silico and then on the bottom are steps that are performed in the wet lab.
16:24
So we start with our parental antibody, and we apply our XtalFold algorithm to create a structural model of the antibody-antigen complex.
16:38
And I'll talk more about XtalFold in just a minute.
16:40
And by doing this structural modelling, just based on the primary sequences, we were able to identify 25 residues that are at the antigen-antibody interface.
16:52
And so from there, we can use our generative AI to create a panel of mutants where those specific 25 residues are mutated.
17:06
Then from there, we can narrow it down to mutants that we consider fit that are recommended based on their developability and then rank them according to their predicted affinity using AI algorithms.
17:22
And with all of those filters applied, we end up with 23 sequences that we actually expressed and tested in the lab.
17:32
And we did this quite rapidly using a cell-free expression system and then used BLI for affinity measurements.
17:39
And based on that data, we fed it back into our AI algorithms for fine tuning of the affinity.
17:49
And from that analysis, we determined that there were eleven single-point mutations that were predicted to be most important for mediating affinity.
17:58
So then we were able to create a new panel of variants and rank them using our affinity AI algorithm.
18:04
And from there, we chose 50 to express and test in the lab.
18:10
And at the end we were able to identify our top clone which had an affinity of 1.53 nanomolar with just five point mutations and that resulted in a hundredfold increase in affinity.
18:24
So this whole process took only three and a half weeks.
18:31
Now let me talk a little bit more about XtalFold.
18:33
So XtalFold, as I mentioned was what we used in the first step of that process.
18:38
And we consider this the crown jewel of our AI platforms at Ailux.
18:43
And what XtalFold does is, it's able to create protein-protein complex structures using just the primary sequence alone.
18:55
So you just feed the model your sequences, and the output is a predicted complex structure. And so you're probably asking how does XtalFold compared to AlphaFold?
19:09
Because AlphaFold is considered state-of-the-art in this field.
19:12
And so we have done a head to head comparison of our XtalFold modelling with AlphaFold-Multimer.
19:20
And the way we did this was we took 39 PDB structures that were in the PDB database but were not used to train either XtalFold or AlphaFold.
19:30
And that served as the ground truth for our comparison.
19:33
So we ran those sequences through both XtalFold and AlphaFold and overall, XtalFold outperformed AlphaFold in by quite a bit.
19:45
So a couple of examples I'm showing you here.
19:49
On the left you have one example.
19:53
The PDB structures are on the top, XtalFold structures in the middle, and AlphaFold on the bottom.
19:58
And you can see for the for the case on the left and the second example in the middle.
20:03
The XtalFold structures look very close to the actual PDB structure, while the epitope and orientation for AlphaFold are quite off.
20:16
In the third example, if you look at first glance, it looks like both XtalFold and AlphaFold are predicting the complex structure quite well.
20:23
But if you zoom into the binding interface, you can see that the prediction for AlphaFold is quite off.
20:30
And so we applied some metrics to really measure the difference in performance here.
20:36
And so what we did with the 39 structures that we analysed, we set a DockQ threshold score of 0.23 and that's the threshold that’s used in the technical community here.
20:51
And so with that threshold, AlphaFold only predicted accurate structures 36% of the time, while XtalFold had a 70% success rate.
21:03
And if we are to apply an internal metric that we've developed called structural confidence, with that score filtration, we then improve the success rate to 90%.
21:16
So XtalFold consistently outperforms AlphaFold in these structural predictions.
21:25
So we have XtalFold as I just mentioned, but also with this affinity maturation workflow, we can apply other AI models.
21:32
So our generative AI model is called XenProt and that is our proprietary large language model, and it is able to recommend developable mutants that will maintain binding to our target.
21:48
And then later on in that workflow, we apply our affinity AI which is called Xffinity, and we have validated this AI model by doing the in silico affinity predictions and then comparing them to the actual measured affinity in the wet lab.
22:05
And overall, that's what I'm showing you here.
22:08
And overall, you can see there's a very nice correlation between our affinity predictions by AI and the actual measured affinities.
22:20
And as I mentioned with our other case study, we can also look at developability as we're doing our affinity maturation and hit selection.
22:33
And so this top clone that I mentioned earlier with a low nanomolar affinity, we were able to also maintain its biophysical properties and even improve them slightly in some cases.
22:45
So we're not losing other desirable properties of the molecule by increasing the affinity of the VHH.
22:56
So to recap here in terms of relevance, we all know that affinity maturation is regularly needed and most often times you don't have structural data of your antibody.
23:08
And so I think this is where AI can be of real use.
23:15
In terms of replaceability, yes, there are existing methods to do this type of affinity maturation, but mutagenesis via phage display for example, is very time consuming and tedious.
23:28
And if you want to apply a rational design approach, of course you need some sort of structure available.
23:34
And so I think in terms of antibody engineering and specifically affinity maturation, this is where AI really shines.
23:44
And we're able to do this affinity maturation in less than a month, whereas display methods might take six months or more to achieve.
23:54
So I think this is really valuable.
23:58
And then in terms of reproducibility, we have been able to apply this workflow to other molecules.
24:05
Here we had a Fab that already had a low nanomolar affinity.
24:10
But after two rounds of optimization and only 5 weeks with expressing 60 variants, we are able to achieve a sub nanomolar affinity.
24:19
So we have been able to reproduce these results for different antibodies.
24:26
And finally, I want to move to our last case study, which involves humanization.
24:32
So here we had an interesting challenge.
24:35
We had a functional murine antibody.
24:38
It was our best performer in our in vitro functional assay, and it was our top lead candidate.
24:43
But of course it required humanization.
24:45
And so we tried the traditional approach of CDR grafting, and we looked at human frameworks that were most similar to the murine frameworks and we also included some back mutations and ended up with 35 mutants total that we tested.
25:10
And so these CDR grafted mutants, when we looked at their affinity and function in vitro, we saw a significant loss in binding affinity and basically a total loss in functional activity.
25:24
So here our standard humanization approach by CDR grafting failed.
25:29
And this led us to consider an AI approach.
25:38
So we started off by generating a predicted antibody antigen complex structure using XtalFold.
25:47
And we found something very interesting when we did that.
25:51
It revealed that there was a specific residue, this Y67 in the framework 3 of the light chain that appeared to be critical for antibody binding and was part of the paratope.
26:03
And so of course we wouldn't have identified that with more traditional methods.
26:09
And so once we perform this back mutation of Y67 on our CDR grafted mutants, we are immediately able to restore a protein binding which you can see on the left with the orange clones.
26:27
And then we also tested them and found target expressing cells. And in terms of function, they were right on par with our parental antibody.
26:38
So I think this is a is a unique case, but very interesting in in terms of how we can use XtalFold.
26:45
But we weren't satisfied there.
26:47
We wanted to take it another step.
26:49
And so what we decided to do was apply what we call our ‘super humanization workflow’.
26:56
And so again, this is guided by AI analysis.
27:00
So we start with the antibody-antigen complex structure, but instead of just grafting frameworks that are most homologous to the murine by sequence comparison, we use our proprietary AI algorithms to select frameworks that not only are similar in terms of sequence identity, but best represent the conformational structure of the mouse antibody.
27:30
And so that may not necessarily be the one that is most similar sequence-wise when you just look at homology.
27:38
And so by applying that AI analysis, we were able to select 4 frameworks that we thought would be the best.
27:46
And then we wanted to take it an additional step further and minimise the total mouse residues in the paratope and CDR.
27:58
And so again, using this type of analysis, we were able to identify the residues that are critical for binding and stability and those that we could vary.
28:07
And so when we did this, we were able to significantly decrease the number of murine residues in the super humanized antibodies which are shown here on the left.
28:19
And then we also did an in silico prediction of ADA.
28:24
And what we found was that our super humanized leads in terms of predicted immunogenicity are as good and if not better than many of the fully human antibodies that are either approved or in the clinic.
28:39
And that's what's shown here on the left.
28:42
So to just wrap up, traditional humanization usually works, but we do have exceptions like what I just showed you.
28:51
So it can be very valuable.
28:54
And XtalFold has been very consistent in terms of the structures that it produces.
29:02
So just to wrap up, AI is more useful in some cases than others.
29:07
Sometimes we find those cases empirically, but we strongly believe in using AI in synergy with wet lab capabilities to increase our chances of success.
29:18
And that's why at Ailux, we have both our computational scientists and our bench scientists all under the same roof.
29:26
So if you'd like to learn more about our company, we have the dedicated biologics brand of XtalPi, which has been around for 10 years, but Ailux just launched this year in early 2024.
29:42
We've worked with lots of different pharma and biotech partners, and we have a whole suite of platforms available for antibody discovery.
29:49
So if you'd like to learn more about any of these platforms are the ones I described today, please stop by our booth #3 myself and my colleagues will be there for the rest of the conference.
29:58
So thank you very much for your.
