0:20
So I would like to break with tradition by starting off with a little bit of doom and gloom.
0:28
If you spend your career at the intersection of science and business like I do, you might see a lot of these sorts of charts across the internet.
0:36
The basic story is that at this time, it's not a really great time to find new funding for early stage biologics discovery programmes.
0:45
There's also a lot of interest from the deal makers in technologies and projects that incorporate some kind of AI.
0:52
And depending on your perspective, that may be good news or may be bad news, but I think it's definitely something that's worth thinking about and talking about.
1:03
I want to preface my talk by saying this is not going to be an AI presentation.
1:06
I'm not going to say, hey, AI is here to solve all your problems.
1:10
I'm a bit of an AI pessimist, I have to admit.
1:12
I do think that at this stage in 2024, there are very specific applications for AI, very specific ways that we can use it.
1:20
And I'm also a mass spectrometry nerd.
1:23
So I'm going to talk about using AI and mass spectrometry together.
1:27
This sad little robot is an image that I actually used AI to generate, which is kind of meta.
1:33
And so I hope the robot overlords will forgive me for that later on.
1:38
OK, so that brings me to my agenda for today.
1:42
I am going to talk about some new technology.
1:44
I am going to put that into the context of some discovery stories and case studies that we've done.
1:50
But I'm going to make sure to address some of the key challenges that I think everybody in this room should be thinking about, which are #1: how to find new things, how to breakthrough some of those deadlocks; 2: how to de-risk antibody discovery, how to get good results while spending less time and effort.
2:09
And 3: how to make things easier for our colleagues downstream who have to produce the darn things.
2:16
So that's the basic outline.
2:19
And I will start with an antibody discovery story.
2:24
So over here on the left … I'll say this is not my story, but it's a fairly typical one.
2:29
This goes back a couple of years to COVID-era. And at the top there is very impressive work.
2:35
This was starting with convalescent serum from a number of patients.
2:39
We've got millions of PBMCs that they whittled down and then managed to get sequences for cloning expression, 187 of them through that process.
2:50
Those were expressed, thrown through a battery of rigorous testing to then come up with 9 antibodies that were ACE2 blocking.
3:01
They have that that activity that they wanted, which is fantastic work.
3:05
I'm not trying to knock it.
3:06
In fact, I'm celebrating.
3:07
It's great work.
3:08
But the problem is if we put it back into our context of what we talked about in the beginning is that there's a lot of money spent at the end of that stage, and 95% of it is spent on clones that were not ultimately selected.
3:23
What I'd like us to get to is a story that looks more like this: to start off with an even bigger, a vastly larger diversity of antibodies that we're going to look at and through some technology whittle it down to a more manageable number of sequences that we're going to ultimately express.
3:41
And for those to then go on to have a higher success rate, so more like 50% instead of that 5%.
3:51
OK, now I can't tell you the whole story about how we did this because the paper has not yet been published.
3:59
But I can share some of the steps that we took, I can share some of the things that we went through to achieve this kind of result.
4:07
So here's the high level method.
4:13
First off is you start with an immunisation.
4:16
So yes, this is an immunisation based approach.
4:19
In this case we immunised our humans with the SARS-CoV-2 vaccine.
4:26
But in principle and some of the other case studies I'm going to show you, we've done it with other species, chickens, alpacas, rabbits, et cetera.
4:34
For reasons that will become clear, we do need quite a bit of starting material for this.
4:39
So we can't use rats or mice, it has to be a rabbit or bigger.
4:43
The next step is we're going to do some purifications.
4:47
This is something that many people have access to this kind of technology.
4:51
It's not super advanced.
4:53
We're going to use protein A/G to purify, get those IgGs out of the serum.
4:58
We're going to use the antigen itself to affinity purify for the subpopulation of those IgGs that are specific to the antigen.
5:07
And we can do some negative selection as well to make sure it's not binding something off target or things that we don't want.
5:13
The result of this is a pAb, in fact, a set of pAbs.
5:16
You might have many, many pAbs, and then you can do testing on those pAbs, functional testing, use it in your assay, whatever you want to do to make sure that that pAb, before you get into the hard work of sequencing and all that kind of stuff, contains the antibodies that you want to work with.
5:35
Then we do the magic of antibody protein sequencing using mass spectrometry and AI and the result of this is usually around 2 to 50 full length antibody sequences and a collection of a few hundred additional CDRs that you can use.
6:01
The next thing that you're going to do is express and test these.
6:04
Nothing new here.
6:05
In our COVID case that we talked about, it was 12 mAbs, seven of them met our performance criteria for binding and six of those showed neutralisation.
6:17
And the last step is to do further characterization, engineering, optimization.
6:24
And again, this is an area where we're going to apply some of that mass spectrometry and artificial intelligence here.
6:35
So here's the whole workflow, and I've highlighted the two bits that contain that mass spec and AI.
6:41
It stands on its own as a whole workflow, but I want to emphasise that these pieces can be plugged into existing workflows to augment, to help filter, and select for the things that you want.
6:56
I'm aware though that I have not at this point explained how any of this works, so I thought I would get into that so that it doesn't seem like a magic wand.
7:06
I asked AI to come up with me with a an artificially generated image and this is the sceptical dude that it generated for me.
7:16
And then I got carried away and I generated some other images.
7:21
The robots can't be trusted.
7:22
This lady has three arms.
7:26
But let's get into how protein sequencing works.
7:29
So the basic workflow for starting with a monoclonal, let’s start with the easy case, is that way over there on the left you've got the intact protein.
7:39
You're going to separate that out into different fractions.
7:42
You're going to digest those fractions with various different enzymes.
7:46
And then you have these colourful squiggles which represent peptides.
7:50
Those are flown through liquid chromatography and then tandem mass spectrometry to measure the masses of the peptides and fragments of those peptides.
7:59
And you end up with quite a big data set and you have to go through that with the machine learning algorithms for de novo peptide sequencing that our company has generated.
8:08
Those sequences are then assembled because those peptides overlap, it’s possible to generate the full length antibody sequence.
8:18
This is something that we've been doing for over 8 years.
8:21
We've done it over 9000 times.
8:22
This is a very robust workflow.
8:27
Polyclonal antibody sequencing, starting with that pad is a lot harder.
8:32
It's a complex mixture.
8:33
So we actually have to do a lot of separations using various separations chemistry which you probably have available in your lab.
8:44
We're also going to look at the intact molecule.
8:47
We're going to cleave off the Fab region.
8:49
We're going to look at the various subunits and we're going to do that bottom up looking at the peptides.
8:55
And this is again where we have to use some pretty advanced algorithms to piece all of that together and then come up with the full length antibody sequences.
9:08
This is what some of that data looks like.
9:10
Across the top.
9:11
You've got the linear amino acid sequence with the positions shown.
9:16
I blanked it out because it's confidential, sorry, but you can see the annotated variable constant regions, et cetera.
9:23
Each of those colourful lines below represents a peptide that we've pulled out of the mass spec data in support of that sequence.
9:31
And you can see how they all stack up and overlap each other to support the sequence of that.
9:37
If I zoom in on one of those, you can see the spectral data that supports that and we've got a nice series of B&Y ions that add up to that complete sequence of that peptide.
9:52
All right, so that's a bit about how it works.
9:56
So let's talk about how we can apply this sort of technology to our three main challenges that I stated at the beginning.
10:03
So if we're going to try and improve the way we search for and find things, one of the things we might do is consider expanding the search.
10:12
If you want to find something new, look in new places, right?
10:18
So what we'll think about here then is increasing diversity and maybe at some point in the future we might be able to use computers to search through all possible sequences and impossible sequences.
10:33
I don't think we're there yet today.
10:35
So I'd like to just start with looking at the diversity of perhaps other in vivo discovery campaigns.
10:44
So maybe hybridomas or single cell sequencing.
10:46
So you're looking on the order of a few thousands or ten thousand of sequences and humans are pretty bad at visualising numbers.
10:53
So my analogy is if each antibody is a grain of sand, then this represents about a handful of sand.
11:01
The AI generated image has two thumbs.
11:03
Don't trust the robots.
11:06
Here's the diversity of a library based approach and phage display or something like that.
11:11
OK, much more diversity, something on the order of 108 to 1010.
11:18
If you've got a bigger library, tell me. But it's truckloads of sand.
11:22
OK, much, much bigger.
11:26
Natural.
11:27
The natural immune response generates even more diversity than that.
11:31
We've got combinatorial joining of the gene segments.
11:36
You got somatic hypermutation, so the human body or the mammalian immune system is generating something on the order of...
11:43
Well, estimates vary… 1014 to 1018 sequences that are then selected, tested against the antigen, et cetera.
11:53
And the best ones are selected naturally.
11:55
OK, so somewhere on the order of that, if you want to know how that fits into our analogy of grains of sand, that's something on the order of all sand on all the beaches and all the deserts on Earth.
12:05
It is a vast amount. But how do we get into that?
12:09
How do we tap into that diversity?
12:12
So you would be forgiven for assuming that if we're going to do bulk NGS sequencing on PBMCs that we might be tapping into a lot of that diversity.
12:23
I thought this and then my colleague explained to me, and she even gave me this whole slide to explain all of immunology on one slide, which is kind of bonkers, so I won't bother.
12:32
But the takeaway point is that the circulating B cells represent only 2 to 3% of the total B cells.
12:42
So you might have enough in that 2 to 3%.
12:46
So remember that number 3%.
12:49
Does anybody know what this is?
12:53
Of course you don't.
12:54
This is 3% of the Mona Lisa, so you might be missing some of the picture.
13:02
So, let's say that you do manage to capture all of that diversity with the B cells.
13:09
There's still another problem is that B cells are somewhat fragile and it's hard to keep them alive and do functional testing and screening on them before you get to sequencing.
13:19
So what I'm going to propose is that we use protein as the analyte.
13:24
It's a very robust analyte and we can do all sorts of functional testing and filtering and things using the protein.
13:32
Here's an example.
13:35
Our case here was an internal project.
13:38
We do work on multiple myeloma at Rapid Novor, if you weren't aware.
13:41
We needed an antibody for our assay.
13:44
We started off immunising some chickens with a protein of interest and we did the workflow that I've described, affinity purification, protein sequencing using our mass spec technology, which we call REpAb.
14:02
We also did the bulk NGS sequencing of PBMC cells and spleen cells and we wanted to compare them.
14:09
So the result was, yes, we got an antibody that was useful for our assay.
14:12
But what I wanted to share with you today was some of the exploring of the overlap of the diversity and where we're getting things.
14:19
The whole number of NGS sequences that we got wouldn't fit on this slide.
14:26
The circle would be too big.
14:28
So I've zoomed in on those sequences where we had CDRs evident in the proteomics data.
14:35
So, of the PBMCs you can see there in in blue, the spleen ones are in green, and I think that's pink.
14:46
I'm a bit colour blind, but the bottom circle represents the actual full length de novo sequences that we found from the proteomics.
14:54
The others are CDRs that we pulled out of the PBMCs and spleen, respectively that appeared also in the proteomics data.
15:02
So the first thing you might notice is that, yeah, there's some good overlap between that.
15:06
There's some good evidence that we can use to say, hey, of those PBMCs, these ones are important.
15:11
They showed up abundantly in the proteomics.
15:14
We can also see that there's a significant number that did not appear in any of the genomic information that we were able to pull out from the protein.
15:23
And we did a cladogram as well to see how these sequences are all related.
15:27
And we can see that, yeah, there is some relation between them and they're also some distantly related sequences that we found with protein only.
15:37
One last thing on diversity is about which animal you're using.
15:42
I casually mentioned that we worked with chickens.
15:47
The ability to work in different animals will give you a diverse immune response.
15:53
Let's look at humans for a second.
15:56
So this is from immunising an alpaca.
16:01
And you can see the graph represents the germ line usage across the bottom.
16:07
And the number of protein sequences identified represents the vertical lines.
16:11
You can see there's three that really stand out from that.
16:16
And here's human.
16:18
OK, so this was from, if I remember correctly, this is from convalescence serum.
16:24
And patient 1 is in red, patient 2 is in blue.
16:28
And you can see that there's quite a few more that stand out above the background.
16:33
And, humans walk around and experience lots of insults to our immune systems all the time.
16:39
So we generate quite a bit more diversity.
16:42
OK, that's the last comment on diversity.
16:45
Let's say we've solved the diversity problem.
16:48
Now we have to think about how we are going to pull things out of all that diversity.
16:53
So let's say you've got your huge planet full of sands.
16:57
You're going to need an interplanetary scale magnet to sort of pull out the ones that you want.
17:01
You don't want to go through each grain of sand of with tweezers.
17:03
So what is that search strategy that we're going to use?
17:06
How do we get from that 1018 down to a few without spending so much money?
17:12
You need the search strategy to select for the things that are interesting: affinity, the epitope that's interesting, function, cross reactivity, all these things and it needs to be economical to do.
17:27
We need some magic glasses to look at that.
17:29
So again, I'm going to suggest thinking about looking at the protein and this is starting at the end.
17:37
This is really hard to sort of wrap our heads around if we're used to doing it the other way.
17:42
So start with a polyclonal antibody that you've prepared, that you've tested, that you know does what you want in your application, and then seek to sequence and understand those antibodies that are in that mixture.
17:57
OK, so I'll give an example.
18:01
This was an antibody drug, antibody project that we were tasked with.
18:06
So I can't give the details of what all that was, but it's a drug and there was another, we wanted to make sure there wasn't binding off target as well.
18:18
OK, so this is by the way, where I completely lose my family in explaining what I do.
18:22
Hang on, you're finding antibodies against antibodies?
18:26
So here we go.
18:27
So we wanted to do this.
18:28
And the way we approach this, is using that same workflow that I described.
18:34
So we're going to immunise a rabbit.
18:35
We've done that.
18:37
We've used protein A/G to purify the IgGs out of that mixture.
18:42
The next step is to do a negative selection using that very similar antibody to the drug to make sure that we're not getting things that are binding to that one.
18:53
So the flow through from that goes into a secondary purification step and now we're selecting for the antibodies that bind to our drug.
19:00
OK, so the result is a polyclonal antibody that works well.
19:05
And we then proceed with mass spec plus AI based sequencing.
19:09
And we generated 30 sequences.
19:14
We chose to recombinantly express 18 of those and of the 18, 16 of them met our performance criteria for binding. Here's some ELISA data to show how well they’re bound – some bind better than others and that's great.
19:31
So there's 2 plots, 8 on each of them.
19:34
It was too hard to put 16 all on one and there's a control which is down to the bottom line.
19:42
Importantly, here's the ELISA against the off target binding.
19:46
None of the antibodies bound the other constant region that we didn't want it to bind to.
19:54
OK, I think I'm running short of time, so I better move on to solving the third problem, which is trying to inform the choices of our colleagues downstream.
20:03
And how do we do that?
20:06
And this gets into the characterization question.
20:10
There are lots of things that we can do to characterise antibodies towards the end of an antibody discovery and engineering effort.
20:19
The problem is that if we're up in the earlier stages, we've got a lot of stuff to work through and it's not economical.
20:27
So we have to go for things that are high throughput.
20:30
When we get down to the later stages, we're doing things that are much more detailed.
20:34
So it doesn't matter what kind of analysis you're talking about; you're going to preferentially choose those things.
20:39
I think our problem as a group is to try and get those detailed analyses to be higher throughput, to give us more information earlier in the process, which will help everybody.
20:58
The here's the mass spectrometry based assays that I can think of.
21:05
So peptide mapping, disulfide analysis, glycosylation analysis, PTMs, epitope mapping, et cetera.
21:12
And AI gives us some opportunities as well: epitope prediction, maybe some sequence liabilities that we can model et cetera.
21:19
OK, so how do we combine those two things together?
21:23
The low hanging fruit I think is on the epitope mapping and the idea is to put these things together to get the robustness and the detail of epitope mapping with mass spectrometry but add in some AI to increase the throughput.
21:42
So if you're not familiar with mass spectrometry based epitope mapping, it's done through a technique called hydrogen deuterium exchange mass spectrometry and the flow is something like this.
21:58
You start with the intact molecule, you introduce deuterium, and the surficial hydrogens get exchanged with deuterium, lock in that by quenching the reaction and then digest the protein, run those peptides through mass spectrometry and then you can observe the mass difference.
22:21
You do this again in the bound state at various time points and then you can measure the differential uptake of deuterium.
22:29
OK, so those peptides that are on the surface will get exchanged except if they're in the binding region, they will be protected.
22:40
So then, that's how you can determine where the epitope is.
22:44
All right, there's a lot to that.
22:47
If you want to know more, come and see me afterwards.
22:50
But it's a great technique and our customers were asking us if we could do this higher throughput.
22:59
Here's what some of the output looks like.
23:01
So the, the blue region is implicated in the epitope binding, and you can see that on a linear basis as well.
23:08
Here's the regions that are implicated.
23:11
This is supposed to represent the amino acid sequence.
23:14
Again, I've had to block it out because it's confidential, but this region clearly indicated in the binding.
23:22
OK, so how can we improve the throughput of this very nice and robust technique?
23:27
So this is what we actually just announced yesterday.
23:31
We announced a partnership with an AI firm called MAbSilico, who's helping us out with this.
23:37
And the workflow is essentially start with say 100 antibody sequences and the target structure, you'll get some epitope predictions out of the AI, essentially some epitope bins generated by the AI.
23:55
Of those, you can choose say 3 bins, 3 epitopes that look functionally or biologically relevant to you, and then do the mass spectrometry based workflow with the HDX to confirm, yes, we have the epitopes here.
24:13
Those experimental data get fed back into the AI engine to do forced docking experiments, which refines the epitope predictions.
24:25
So what it essentially means is that you get in a similar kind of time frame.
24:32
So HDX takes about four weeks, this takes about four to six weeks, and you can get good quality epitope information for 100 mAbs at a time instead of one or two.
24:45
So that was just announced and I’m able to show what this looks like.
24:54
Yeah, there's a 3D representation of those epitope bins, so epitope bins one, two, three, and four and showing the region that's part of the interaction.
25:08
OK, so I'm coming to my conclusion, and I've done this in the reverse order because I haven't introduced my company or myself at the beginning.
25:18
So I'll make sure and do that now.
25:20
So Rapid Novor is a mass spectrometry-based services company.
25:25
We do work in the antibody discovery space.
25:30
We've been around for since 2015.
25:35
And we've done a huge number of projects.
25:38
And if you want to talk to us further, come up to me afterwards or see us at booth 44.
25:44
Here's a list of our services.
25:46
So I mentioned monoclonal antibody sequencing, polyclonal antibody sequencing and antibody discovery, epitope mapping using HDX and we've got capabilities around binding kinetics using SPR as well for screening and binning and profile.
26:02
We also do offer these characterization services, peptide mapping, glycan analysis, all that sort of thing.
26:11
And this is my last slide, and I'd really like to thank these individuals who are behind the technology and the programmes that we've discussed today.
26:22
In particular, Doctor Xiaobing Han provided me with many of the slides and did a lot of the work that's behind what we talked about today.
26:31
Unfortunately, she couldn't be here, but if you enjoyed my talk, perhaps join me in celebrating Dr Han and all of her great work.
26:40
Thank you very much.
