0:07
Nevertheless, just arrived.
0:10
So I will not be able to refer to some of the other talks that has already been taking place about weight lab, dry lab and in silico discovery of antibodies.
0:22
But here is what we'll be talking about today or what I'll be talking about today, just a very high level introduction to the field.
0:32
We'll talk a bit about the challenges that you all have when you're trying to generate therapeutic antibodies and how we as part of Bench Link solve that with data capture analysis and automation as well.
0:47
And then at the end, we'll have time for a few questions, hopefully.
0:52
So bear with me.
0:53
20 minutes and then we have coffee and cake in the hall.
0:58
So we all know that or most of us know at least that 12 of the top 20 best-selling drugs are monoclonal antibodies.
1:11
This has been going up over the last 10-15 years from very few to now more than half.
1:18
And those 12 antibodies generate more than 60% of the total drug revenue here.
1:25
So this is of course super interesting for pharmaceutical companies because this is where there's a lot of value in the market.
1:33
And from my perspective personally, there's of course an interest to actually generate these monoclonal antibodies so we can actually treat some of these vicious diseases.
1:45
This is probably also something that you all know that the time to market is extremely long and is extremely expensive.
1:54
So going from preclinical or discovery where we have millions or billions of data points to actually a final FDA approved antibody, it takes way too long.
2:08
And one of the opportunities we now have is to shorten that time by using AI, more automation, better capturing of data and so on.
2:25
So a bit about the challenges, and I'm sure we all have the same challenges, but just sort of a very brief historical perspective.
2:38
If we look at Pub Med and we just do some basic literature searches, we can see that if we search with antibody as a term and either machine learning or deep learning or other machine learning terms, we'll see a very steep increase in the number of publications.
2:59
Interestingly, the good old phage display number of publications is actually going down.
3:05
So we're shifting from a more wet lab interest to a more dry lab interest in by using machine learning.
3:16
Craig Winter received the Nobel Prize on phage display in 2018 and just last year David Baker and more was awarded the Nobel Prize due to his influence in computational protein design.
3:35
So this is interesting to see the trend not just in the delivery of biologics to the market, but also we can actually see that as well in the literature.
3:47
So the challenges that we have and I guess that's all of you also have is that the amount of data that is being generated is increasing.
4:00
The complexity of the data is also increasing.
4:03
And this is only for the wet lab part of things.
4:11
But if we just look at this very complex mind map of data being generated, we of course have sequencing data.
4:19
That's where PipeBio comes in as a bioinformatics platform.
4:25
But we also have high flow mappings such as LIBRA-seq or BEAM from 10X.
4:30
We have binding with kinetics and so on.
4:34
We have developability and then we have assay data.
4:37
We have public databases as well.
4:40
So all of this comes basically from the dry lab, and we need methods how to capture and automatically capture all of this in the structured manner.
4:54
On top of all of this, as an example, for developability, for aggregation, as an example, we also have computational methods.
5:03
So we have not just only the wet lab results, but we also have this from the dry lab and computational in silico methods.
5:14
So we have a challenge in just how to capture all of this data in a structured way.
5:20
And then on top of this, the 12 antibodies from before were monoclonal antibodies, but now there's also a tendency to make more funky scaffolds on top of the good old IgG also having other modalities such as DARPins, bicyclic peptides and so on.
5:44
So the amount of different scaffolds or modalities being used together with the increased number of data points is very challenging.
5:57
And again, so the high throughput assays and tools for characterization goes across the entire flow all the way from the synthetic libraries and animal immunisation all the way to in vivo evaluation and optimization.
6:15
And you will see this throughout this presentation that there's unfortunately no linear way through antibody discovery, at least when you have AI involved.
6:28
There's a lot of looping back and looking at data again, generating new models and so on.
6:35
And this will also be what you'll see here in the presentation that it seems a bit sort of disrupted at times because we then loop back to some other methods just to illustrate some of the different modalities that we are challenged with.
6:56
So here on the right-hand side from sort of a Benchling, we have of course the IdGs, but we also have scfvs, fabs, nano bodies and so on.
7:08
And this is all what is needs to be registered in the database in a very structured manner.
7:14
Because when you have registered all of these your variable domains and whatnot, then you can start building your bispecifics or more advanced complexes and scaffolds.
7:33
So going back to the complexes and the scaffolds that's where it all starts.
7:43
And it's very important to have that registered in a very structured manner.
7:48
So when I wear the Benchling hat, we have this structure data in a database and then that can be enabled through high throughput workflows.
8:02
And from the high throughput workflows, we of course can automatically capture and analyse the data.
8:08
And then we move into PipeBio to look at the sequences again and overlaying the structural and functional data.
8:14
And then we sort of cycle around until we actually find the antibody of interest.
8:22
So just stepping back, if you're not aware of what PipeBio is bio is, it is the bioinformatics cloud for scientists.
8:30
So we focus very much on very large-scale sequence analysis of antibodies, primarily NGS sequencing but also from Sanger sequencing and everything is in one integrated platform.
8:44
But now as we are also a Benchling company, we have the entire end to end flow from raw sequence data to hit picked and databased and registered and so on, even for these modalities that you've seen on some of the other slides.
9:05
So we go from sequence analysis to protein engineering, expression and purification, in vitro and in vivo testing.
9:12
All of this is captured in one platform.
9:18
Just going back to some of the methods that we're using for developability prediction.
9:27
Unfortunately there's no single metric just yet at least to detect whether an antibody is developable.
9:38
But these are also some of the things that we are working on: more stability, thermodynamics, structure, surface exposure and all of that both for sequence-based liabilities, but also structural based liabilities.
9:59
So just to extend a bit on some of our developability prediction tools in the PipeBio platform.
10:07
So of course to predict developability, one often would like to have a structure prediction and this is what we have here.
10:17
On the left-hand side we have the structure of an antibody and we have a linear sequence where we can actually see aggregation prone regions.
10:27
Here on the right-hand side we have a few developability charts or metrics where you can see the all the dots indicate individual antibodies and in high throughput we can analyse sequences and show where they will actually fit into a developability sort of scale or metrics.
10:50
So the purple violin plot is basically a background distribution of known therapeutic antibodies.
11:00
So then we can see do our new novel antibody from a repertoire sequencing or something similar, will that actually fall into the same sort of category or placement as some of the therapeutic antibodies?
11:15
What we have done is to make this incredibly scalable.
11:19
So not just predictive data from PipeBio or from Benchling, but also from predictive data from your own sequences.
11:28
So you can, with the tool build your own models in the dry lab.
11:34
You can run or use that data to actually generate a background distribution of your already qualified data.
11:44
And then going forward, when you do repertoire sequence or sequencing or campaigns, you can actually compare to your already existing data to see if it sort of falls within the threshold that you want.
12:00
So all of that is high level shown here.
12:05
So we have on the left hand side, the wet lab where we have instruments, we have scientists and systems.
12:12
And here that is all captured in Benchling.
12:16
And then here on the right-hand side, we can use all of that data or you can use all of that data to generate your own models and then bring that back in and register that with your ready existing data.
12:35
We of course also have visualisation.
12:37
So when you bring in your new model data or from already existing tools in the platforms, these data points can be overlaid with some of the existing graphics.
12:48
So if you have built your own machine learning model to predict whatever, you can bring that data in and you can use that to overlay on some of the building charts in the system.
13:02
Hands on, you will do sequence analysis and hit picking alignments and so on.
13:09
With the sequences in PipeBio, you'll select your hitpicked sequences, you will push it to Benchling and they immediately be registered there and that will close off the loop of the data registration.
13:23
This is along the lines of what it'll look like in Benchling that you actually have a registered sequence, you'll have your amino acid sequence and then you'll have information of metadata on your antibodies which you have just pushed into the registry just to show a recording of some of these high throughput assays.
13:45
So here we just have a microtitre plate to visualise.
13:48
And then again, you can do your assay registrations or data outside in the web lab, bring that back in and again enrich your already existing data in the platform.
14:06
So again, just to reiterate it's important to capture all the data.
14:19
So we have a lot of built connectors where you can automatically capture the data from your instrument.
14:25
So you don't need to export data from your instrument and upload manually to the platform.
14:30
So all of this data can be captured and registered automatically.
14:35
And then you can basically focus on the science regardless whether that's a web lab user or a dry lab user.
14:48
So as a scientist, of course you want to also view the data.
14:53
And now we're challenged because now we have a lot of different data types.
14:58
We have sequence data, we have functional data, we have structural data and we have these capabilities of making dashboards.
15:06
So you can actually view what you're interested in.
15:10
And this is of course different groups and different research areas have different interests for these dashboards.
15:19
And of course this is complex to write SQL queries to actually to generate these dashboards.
15:29
So here's the SQL query.
15:34
I bet that most wet lab users will not be able to write this, so we developed a large language model to actually help you out.
15:41
So for the web lab user in particular, we have made it possible to just type in what you actually want to get out of the data.
15:50
We will have give it our best shot.
15:55
Propose the SQL syntax that you actually need to write and then you can have your dashboard easy as that.
16:03
We can do this for the dry lab scientists.
16:08
They probably can figure out to write the SQL and use the API to get the data out if they want to.
16:13
But this is also built for the wet lab scientists so they can actually look at the data and not needing to write SQL queries.
16:24
So they can just basically type in what they want to get out of the data.
16:29
Super nice.
16:36
Yeah, I'll just skip this basically.
16:41
But again, a full platform with your sequence or with your sequence analysis as well.
16:47
So when it's brought in your hit picks, you can also clone them and visualise them here.
16:51
So you don't need to jump to another desktop application to actually do most of your work.
17:00
So bringing it all together, we basically want as scientists to actually be able to twist the knobs with aid from the wet lab scientist, the dry lab scientist.
17:16
And bringing it all together is basically around what we just mentioned.
17:20
It's the data capture; it's the powerful sequence analysis with structural predictions and developability assessment.
17:29
And then for the dry lab to actually deploy those models, bring in the functional data or the predicted data from the models and then basically iterate until you find your target lead.
17:45
And with that, I just want to show the capabilities of the Benchling product offering.
17:52
We cover everything from very early states all the way down to process development and in view testing and what not.
18:03
So with that, I'm a bit more relaxed now, so I'll take questions.
18:09
Thank you very much.
