0:00
Welcome everyone to our workshop.
0:03
So we're going to be talking about really how computational and methods can be used to really help augment your drug discovery pipeline and really make better molecules faster.
0:16
With me today, I have Esam Abulrous.
0:18
I'm Ilaria Salutari and myself Dan Cannon.
0:20
We are in the application science team at Schoninger.
0:23
With us also is Joseph Mulvaney.
0:25
He is from our Enterprise Informatics platform department and we'll be talking about that towards the end of the presentation.
0:33
And so this presentation is going to take you on a bit of a journey.
0:36
And again, if you're very familiar with modelling or even if you're not familiar with modelling, if you don't do modelling at all, or you want to know about data management in the way that you can really leverage the data that you have to enhance your drug discovery process, all of that will be covered.
0:52
Now the session, we wanted to be quite interactive.
0:55
Please feel free to raise your hand and interrupt if you have questions.
0:59
We'll try and keep things on time.
1:00
I may might defer to the break for some questions.
1:05
And so we'll just start by introducing Section 1.
1:07
We'll really be on the protein modelling side of things and looking at the technologies we have there to address some of the most common questions that you will have in your discovery process.
1:18
We'll have a break where you can ask questions and then we'll talk about, yeah, our enterprise informatics platform, LiveDesign and really how we can use that through a few examples on how to, again, leverage that data.
1:32
So within Schrodinger, you know, we do have a sort of certain vision for biologics discovery where of course you want to ideate and take quite a lot of examples.
1:44
But what we say is that if you can actually accurately calculate properties of your antibodies, you can in theory bring the single best molecule to market.
1:55
And so this is a molecule which is meets the criteria of being safe, efficacious, developable.
2:01
And this includes many things such as stability, affinity, selectivity, solubility, immunogenicity, viscosity, manufacturability, all of these things are important factors.
2:10
And we do have aspects which can enhance and help you rationalise the behaviour of your molecules and actually make better decisions.
2:19
And so you hear a lot of course about using machine learning and especially AI models.
2:26
And this is something where we do leverage the advantages of these models along with what our core technology is.
2:34
And this is physics based methods.
2:36
And so just a quick summary, machine learning and artificial intelligence methods are of course very good at interpolation.
2:43
So if you show it, I mean the old, if you show it many pictures of cats and you show it a picture of a cat, it will tell you that it's a cat.
2:50
But it does have some challenges, you know, so it cannot extrapolate in terms of things that it hasn't seen before.
2:57
It's obviously very fast and can handle very large data sets, indeed thrives on large data sets, but of course it requires that.
3:06
Physics based methods on the other hand are they don't require any specific training set.
3:11
These are explicit calculations of physical phenomena and these can be advantageous in that you can extrapolate into novel chemical space something that you might not have tried, you have an idea in your head but would take a long time to make experimentally.
3:26
You can try that in silico and actually fill in these sort of the dark spaces of your chemical or mutagenesis or sequence space.
3:35
They're accurate, but they are, you know, tend to be slow, especially when you go to combinatorial explosion.
3:41
And so this is where we can augment these together.
3:44
So use, you know, take our physics based methods and apply machine learning and AI methods to those to really get to larger scales and to, you know, these dark places within our sequence space that we can, you know, that we wouldn't be able to extrapolate otherwise.
4:03
And so bringing all of these together.
4:05
And so that's just the start of the introduction.
4:08
And so I'm going to hand over to my colleague Esam to take you through some of the backgrounds and some demonstrations on our protein structure prediction properties.
4:18
Thanks.
4:19
Hi, everyone.
4:22
OK, so to engage with Schrodinger tools, we have two wings, two kinds of ways that you can engage with Schrodinger.
4:32
I'm going to cover the first part, what we call Core Suite.
4:36
This is software that you install and by this software then you have access to multiple tools that will help you to perform your calculations the way our tools help you.
4:50
Generally, I'd like, I always like to look at it as different steps, different levels.
4:56
And this matches with your starting point, because in biology, sometimes you start with only a sequence with some experimental data.
5:05
Sometimes you have a sequence and you have the structure, but you want to understand where are the structure.
5:11
Sometimes you have the sequence and the structure and you have some information, but you want further deeply understand the structure, the liability issues that might be there.
5:23
And sometimes also you have the structure you under understand your system, but then you want to engineer a better variant of your system.
5:32
So for each of the of these levels, you have a bunch of tools that will allow you to run those calculations in silico before actually you go to the wet lab and run your experiments.
5:47
Let's start with one of the very first scenario, I just have a sequence and I would like to maybe understand how this sequence arranged in 3D.
5:57
Historically this usually was done with homology modelling and I believe you probably are familiar with this homology modelling.
6:06
In order to have a good model then you need basically a good template that basically with high similarity of your query sequence and then you build your model based on this template.
6:21
The more the better the template is and the more the template is similar to your query sequence, the better the model at the end.
6:29
And this is part of the tools that basically is available in Schrodinger.
6:38
Specifically for antibodies, this is a special case because we use this template way of like you find a template and then you build your model based on it.
6:46
But basically we do this for antibodies in stepwise manner because we do this for the framework, then we do it for each of the CDRs and then we grab the CDRs into the framework and then you just minimise and make sure that you optimise your resulting model.
7:02
This is good but it takes time as well.
7:06
So you need a few minutes to run this and you get a model not very optimal if you have a long list of sequences and you want quickly to build models for.
7:18
And that's why we also now adopted the immune builder tool within our suite and basically use this in our back end and build on this to allow you to build a full antibody refined using the prime like basically our tools to at the end get a full antibody or maybe Fv or Feb structural model refined and optimised.
7:45
So then you can actually work on it and perform calculations.
7:52
And just to give you a glimpse of how this would look like in general via the graphical interface, each type of calculation you have accessed by a panel.
8:01
So structure prediction panel would look like this.
8:05
And basically this is the panel used for homology modelling.
8:08
And now we have actually basically a tick here that basically would enable using the AI or machine learning based models to build a model of your antibody.
8:21
Good.
8:22
So now one scenario I have a sequence, I want to build a structure.
8:28
So another scenario would be I have a structure, crystal structure.
8:32
Usually crystal structures have issues.
8:35
And what you would like to do is OK, I have a structure.
8:38
I would like actually to understand what are the structural issues, maybe strict clashes, missing atoms, missing loops.
8:43
And I want to do this quickly.
8:44
I can also I can do it in a batch mode.
8:47
So I have several structures and I want to run through them even with the model that we discussed, like if you generate a model, you want to also look into the model.
8:54
If the resulting model, especially the one with machine learning based approach, sometimes you have clashes, you have issues.
9:01
They're the don't make sense.
9:03
And for that we have a bunch of panels or a bunch of tools that will allow you to do this by just as easy as just one click.
9:10
So one of them will be the structure reliability report.
9:12
And this is just a panel.
9:14
You import your structure, it analyses the structure and then gives you back a list of different like a a list of assessment of different issues.
9:25
If there are missing loops, missing atoms, steric clashes and so on, they should show up all here.
9:32
OK, we did this.
9:33
We know that I have some steric clashes, I have missing atoms, missing loops and maybe some improper dihedrons and so on.
9:40
I want to deal with this.
9:41
I want to resolve this so quickly.
9:42
I can use this optimised structure for structure based calculations.
9:47
And here we have another panel that basically again, we optimise this with internal benchmarks that basically then you just need to import the structure.
9:56
It's going to go through the structure in a stepwise manner and resolve all these issues and return back an output of a structure optimised, refined and minimised.
10:07
Sounds good.
10:09
Just again give you a glimpse of this.
10:11
This is how it looks like.
10:12
You import the structure and as I mentioned three steps of preparation and then you the result will be an optimised structure.
10:20
OK, I have an optimised structure.
10:23
Now what can I do with it?
10:24
There are different scenarios. You might be interested in developability.
10:29
I want to understand if my structure has some prone like developability issues, may be prone to aggregation, maybe have some exposed reactive residues, exposed cysteines that might result in maybe dimerization and so on.
10:44
Another question would be like, OK, I understand all of this.
10:47
I want to engineer over a variant.
10:49
So I want to study impact of mutations on stability, affinity and so on of the structure.
10:54
Different question would be protein-protein interactions.
10:58
Another question would be dynamics.
11:00
For example, for an antibody, I want to understand how dynamic those CDRs together and so on.
11:05
For today's workshop, I'm going to focus on the first two questions.
11:12
Sort of I'm going to.
11:13
So it's going to be a mix of like I show a slide, I discuss and then I show you how this is done in the software.
11:21
For developability, we have actually many tools that would help you to assist the developability of your input structure.
11:32
One of the most actually famous parameters that we calculate to predict if a structure is prone to aggregation or not is what we call AggScore.
11:47
So AggScore is a numeric value that at the end would tell you if a structure is prone for aggregation or not.
11:56
And the way this goes, it runs.
11:58
So it takes the structure analysis the surface of the structure then identifies patches and then tell you like this patch is with this size is positively charged, negatively charged, or prone to aggregation.
12:10
So calculates aggregation propensity using the equation shown here to show you how AggScore as a value performs in general.
12:23
This is a case study that we did internally.
12:28
The MEDI 1912 and MEDI 578, those, I mean they are all published here.
12:33
MEDI 1912 is a high affine [unclear] variant of MEDI 578 that was found using phage display.
12:44
It's tenfold more affine.
12:48
But the problem is, I mean this is this has been done like basically they found like 12 different multi mutations.
12:54
But the problem is it's like more affine, but those mutations introduced aggregation propensity or like some aggregation problems into the system.
13:04
And you see this from the chromatogram here.
13:07
We took those like we took 578, MEDI 578 and MEDI 1912.
13:12
And then we ran aggregation like the surface analyzer or aggregation propensity predictions.
13:17
And we see on the left hand side MEDI 578, right hand side MEDI 1912.
13:22
And you see actually that this is represented.
13:24
Now this red badge is the one responsible for the aggregation that have been seen experimentally.
13:32
They also identified that three amino acids that were mutated are responsible for this aggregation.
13:38
And I'll show you in a second that we could also see this from the panel. Back mutation helps and they in the paper performed several multiple mutations there.
13:52
And you could see that we took each of them and calculated export for each of them.
13:56
And this is how the correlation between the predicted export and the retention time would look like.
14:03
I mean, so how do, how easy to perform this?
14:07
Like is it like complicated or can I just like have a structure and just submit it and can, how can I interpret the data and the results?
14:15
So now I'm going to switch to basically our software like BioLuminate.
14:21
This is how it looks like, as I said, like it's, you know, a window that gives you access to many tools.
14:28
Good.
14:29
So I have on the left hand side here MEDI 578 structure that looks like this and I have MEDI 1912.
14:40
And as you see from the name, it's prepared.
14:41
So I already did the job like I already prepared the structure, I made sure that everything is fine and so on.
14:48
So now we know that MEDI 1912, I mean we know that's the aggregation prone variant.
14:56
So how could I predict in a way aggregation propensity or like calculate AggScore there?
15:03
So let's protein surface analyser.
15:07
I mean, I ran the analysis a priori, but like I'm going to show you in a second, bear with me.
15:17
Yeah.
15:18
So ideally this would be blank at the beginning and all what you need to do.
15:22
So here this table would be blank.
15:24
And then all what you need to do is just hit analyse.
15:27
It's going to pick the structure from the workspace here that I have in the workspace and then calculates all those parameters.
15:34
OK, how this is going to look like once the analysis is done, this is how it's going to be represented.
15:41
You see this is the batch that had been identified.
15:45
OK, how does it look like in comparison with 578?
15:50
This is how 578 look like.
15:51
OK.
15:52
Again, I calculated the AggScore values.
15:55
So as you see there is a batch that's much smaller in case of 578, but it's much more pronounced in 1912.
16:04
This is the representation on the workspace.
16:07
But then do I get more analysis in there?
16:12
Yes.
16:12
So once you analyse you get back this.
16:15
And as you see, one thing would be first patch browser like it identifies different patches surface patches on the surface of the protein and gives you more information about each batch positively charged.
16:29
And if we Scroll down negatively charged and if we Scroll down hydrophobic, for each one of them, you get the size of the patch, the score and so on, aggregation information.
16:41
So you get again information about what part, what residue.
16:45
And then you get not only the AggScore that I showed on the slides, but you get actually more scores that like, I mean, common in the field.
16:55
So you some other aggregation propensity parameters like scores would be also calculated, but the AggScore value would be here.
17:05
OK.
17:06
And if I click here and for example, I can just sort them based on this.
17:10
And you see now that I have some aggregation patches or that show up here and those are the ones that are highlighted here as well.
17:22
OK, Any questions so far?
17:30
Yes.
17:31
So there is a crystal structure of 578 that is published and then 1912 was generated from it.
17:38
So the structure is available for 578, yes.
18:00
When you generate the structure and it might say it identifies the areas that need correcting and then you say it fills them all in, how good is it at that?
18:09
Does it give you an indication of how successful it was or in building the model?
18:13
Yeah, Yes, I mean it in general it's good in catching.
18:18
So there are two things.
18:20
Number one, machine learning based or homology modelling.
18:23
In terms of homology modelling as I showed like if you have a good template with high similarity, so you have much higher confidence in the resulting model, that's one thing.
18:32
Machine learning, alpha fold and all these now available tools.
18:37
I think the model quality is good especially in the backbone fold.
18:42
Like catching where the backbone would be positioned.
18:46
One has to maybe then think about maybe generating several to catch more information about the side chain packing that might like.
18:56
Generally I would advise generating several models even for homology modelling.
19:06
Yes, I mean for homology modelling you get the reliability report that I showed.
19:11
It gives you confidence like what issues showed up so you can have you can assess, ah, I need to deal with this, I need to deal with that.
19:20
When you can apply this also with machine learning.
19:22
So in general, it takes also a structure as an input.
19:25
Even if you have a structure, not a model, you need to also look into it because crystal structure, for example, you don't have hydrogens.
19:32
So you need also if you have a histidine, if you have aspartic acid like titratable side chains, you need actually to predict how the hydrogens are can be added.
19:42
And this is done within the protein preparation.
19:45
Any other questions, Sir?
19:48
Sorry, I just hello.
19:52
Yeah, so I have two questions, one just related to what was asked right now.
19:56
I'll ask the second one first, the ASCO numbers.
20:00
Is there a reference that number for proteins that you know above which or below which you would consider it a proper aggregation profile or?
20:13
You could eject.
20:14
I wouldn't compare across different structure, but and it's more it's also I will show you in a second.
20:20
It's also not necessarily like if I have one residue that shows high AggScore would be relevant in a way.
20:27
So ideally I would take it case by case in a structure.
20:32
But then for example, in this case, as I will show in a second, you have 3 amino acids that form a batch.
20:39
So we are in 3D proximal to each other and they all have high AggScores.
20:44
If you have such a thing with a batch that is big also in the surface area, that's a concerning.
20:49
That's a signal.
20:51
You could see also like in the structure here, I have some other ones here and there, you know, but this is very pronounced.
20:58
I mean three different amino acids, as I will show in a second, mainly 3 different amino acids that are responsible and they all show high AggScore.
21:08
Sam, I can add a small point to that.
21:09
Is that what we're showing here in this example is the ranking of variants.
21:13
However, of course, looking at diverse antibodies, there was a paper from Charlotte Dean's group looking at therapeutic antibody profiling where they basically came up with a Lipinski’s rule of five for antibody developability related to CDR length, total CDR length, hydrophobic patches, negative patches, positive patches and charge balance.
21:33
Those can actually be used a bit more to compare diverse sequences.
21:36
But there's, you know, it's not as maybe precise, but it's something that can give you a guide is this developable?
21:42
Is there a, you know, the CDRs are quite long, Is it that a flag that I need to consider?
21:47
Is there a patch there?
21:48
So we can do that and we can calculate those parameters as well.
21:52
In general, just like in general, it's a system.
21:55
So you need to generate like in this case the dimensions like there is already a reference not from the calculations like you know you have out there available through UTIC ones that you can run your calculation define your reference and then relate to it.
22:09
You could do this for other systems like as well.
22:12
Yeah, OK.
22:13
And then going to my first question which was earlier, sorry the for the side chain accuracy which he just asked if the backbone is good enough, if you feel its good enough, how much variation is allowed in your this calculation, the cut offs etcetera.
22:31
If the side chains were badly placed, would this be reproducible among is there a systematic study that you have done related to this?
22:40
If the protein structure side chain placement is not great, do you remember any?
22:47
I mean, for me like this is like there will be always.
22:50
So the Rothman's for example, will be.
22:52
So you would be able to change them if you want.
22:55
But in general, I'm not, I don't remember.
22:59
Like if you have a systematic study where we, I don't think we have, I mean we haven't got a systematic study.
23:04
But anecdotally, what we're doing is, yeah, we you can build the model and it will place the side chains in one position, but we then have algorithms that will do cooperative side chain sampling.
23:15
So of course you know it will actually look at this in local environment, but I don't have any benchmark data there.
23:21
OK, So you would suggest that you look at all of those and do the patch on more than one structure.
23:29
You know, if you don't, if you suspect that the side chain placement may be wrong, you do it on a series of structures suggested from the tool that you just mentioned.
23:39
So yeah, I mean looking at the results, if I take the correlation with AggScore, these are with just single output models.
23:45
It's normally not necessary to make multiple models.
23:48
You can get away with one model.
23:49
It doesn't need to be, you know, exactly the right side chain placement.
23:52
It can deal with a little bit of fuzziness.
23:54
OK.
23:55
Sorry that got very long.
23:57
Yeah, no, it's fine.
23:58
Yeah.
23:58
So maybe we move on just now and then.
24:02
So thanks.
24:03
As I was mentioning, like for example, if we look at this list now, we see that like the like a few amino acids that are forming this big batch that is causing this aggregation propensity.
24:16
The in the study they what they did is OK, let's reverse like back mutate those to the original protein and see if we can actually resolve this issue or not. Another that's actually good idea, but like just to demonstrate like how you can do things differently in silico, then OK, one could say, OK, but I can actually maybe run saturation mutagenesis on those 3 amino acids to see if I can actually come up with an idea of different variants that will maintain the tenfold affinity.
24:48
But I would resolve this aggregation problem.
24:52
And this brings me to basically then the second set or like the next set of tools that we can that you can actually use to model or engineer new variants to optimise certain property, in this case stability and affinity.
25:11
In general, what we do this like I'll show you maybe at the end of the presentation, a new features, new tools that we have implemented.
25:18
But what we usually like calculations that we allow or like we offer here is like two MM-GBSA or residue scanning MM-GBSA and protein FEP resid ready scanning MM-GBSA.
25:36
As it's shown here, it's quite fast.
25:39
So you can run a lot of mutations, predict impacts of mutations on affinity or stability.
25:44
And that's because it takes in a rigid structure that it doesn't look into dynamics and it use a simplified model or like approximated model for the solvent.
25:55
Like you know, there is a lot of approximations in there that will allow for a quick qualitative way to assess impacts of mutations.
26:05
We're talking about like one minute on your laptop for mutation and basically then you can imagine now then I can run a lot of those and that's why ideal scenario for this.
26:19
I have many ideas, 102 hundred, 300 and so on.
26:23
And I want just like maybe to identify which ones that I should avoid, like which amino acid, which mutation ideas that I can.
26:30
It's just a no go option for me to you need a complex.
26:41
Yes, you need a complex there.
26:43
So either like you can build a model or you can have the structure of the complex.
26:51
And that's one.
26:51
And that's why we usually recommend using this for enrichment.
26:54
Just which ones I should avoid?
26:57
But then the second tool is like protein FEP.
26:59
This is highly accurate because it used, it's basically the calculations are run on MD simulation.
27:06
So if there is flexibility in the system there, it allows for this and then which allow also for much more accurate prediction for the of for the impact of mutation.
27:16
You have also an explicit representation of the solvent which means like any water mediated interactions will be actually seen there as well.
27:25
But of course then it needs investment in the compute resources.
27:31
I mean, maybe I can tell you that at the end of the presentation, I'll show you now that we have a new method that is in between these two methods, not very highly qualitative, but at the same time, it doesn't need that much resources.
27:46
So you can use it also as a pre filter for your calculation before you go to FEP or to the wetland MM-GBSA again, the same way it's a panel allows you to import the structure.
28:00
And then basically you can run the calculations.
28:03
So this is the panel you import.
28:05
Once you have imported your structure, you have a list of all the amino acids there.
28:08
And then you can introduce the mutation you introduce or like the mutations that you want to calculate.
28:15
And then this is how it's presented, how the results are presented.
28:18
How do we do this?
28:20
Let's go back now the idea in the scenario we are discussing, we have 1912 and we want to maybe run saturation mutagenesis, add the surface where I have those patches in there where I have this aggregation prone patch.
28:34
So if I go back here to the structure and I can just go to MM-GBSA residue scanning calculations, this should be this one.
28:46
And again, I can just import the structure.
28:51
Now I have a list of all the residues, what chain, what residue number and so on.
28:57
So I showed in the slides residue number 30, 31, 56.
29:01
Those are the ones so I can just Scroll down here.
29:05
This is the heavy chain 30, 31.
29:09
It's as easy as I have 30, I click here.
29:12
I have now a list of like mutations that I can introduce in there.
29:17
If I click here select all the natural amino acid and I can even like calculate for the different protonated or like different titratable side chains to protonated or deprotonated aspartic acid, glutamic acid and so on can do this for 30 and 31.
29:40
And as we say, as we saw 56, right?
29:50
OK, I precalculated everything just to save time.
29:56
So it's as easy as just like you can hit run and it's going to regenerate all the mutants and it gives you also an estimation here.
30:03
So like we have now three sites that will be mutated out and then you will have 69 resulting mutants.
30:13
OK, so how do we if I hit run, it's going to calculate all those and then I can maybe show you how the results would look like.
30:25
You get a table that looks like this, the residue that has been mutated, the original amino acid, the mutated amino acid.
30:34
And then you go delta affinity, delta stability and so on.
30:38
And you have a graph that basically shows you, for example, in this case it shows you the delta affinity.
30:43
So the impact of the mutation on the affinity.
30:46
You can of course show the impact of the mutation on the stability.
30:49
And the rest of the table is basically the components because when you introduce a mutations, you impact the column interactions, you van der Waals interactions, you have some desolvation penalties and so on.
30:59
OK, good.
31:01
So if I just maybe run with this scenario and I say, OK, now as I mentioned MM-GBSA, it's qualitative way of assessing the impact of mutation.
31:12
So it's just enough to use it to say which ones are possible or potential good mutations or which ones I should avoid.
31:20
So if I just maybe click here, it's going to basically maybe arrange them this way.
31:26
So in terms of energy, the positive the energy is, the energy value is.
31:31
So this is unfavourable.
31:34
The negative value would mean it's favourable mutation.
31:37
So ideally then if this is qualitative.
31:40
So though this, these ideas might be the ones that are useful because resulting in like negative energy values which means like favourable ones.
31:56
OK.
31:58
I mentioned the 2nd and more accurate way of assessing impacts of mutation would be the FEP.
32:05
You have a handy option that you can just go and send to FEP.
32:08
So it's going to send this list of mutations to be calculated using FEP, which is being calculated through a different panel.
32:16
Before we go to the panel for FEP, let's just maybe understand what FEP is.
32:21
So unlike MM-GBSA, which literally basically if you have, if I have an alanine as a starting point, I would just like immediately switch it to tyrosine tryptophan like in one step right away.
32:36
And that's why it's qualitative in a way here.
32:40
The trick is I'm going to gradually transform this side chain into a new side chain, which allows me then for much more accurate assessment of the impact of the mutation, especially of course if there are some newly water mediated interactions that will be happening around this mutation since it's running also on MD simulation.
33:05
This means then that if there is some adaptation of this introduced mutation like alanine to tyrosine.
33:14
So this is being allowed because the whole system is allowed to move.
33:19
We use this idea of gradually transforming the side chain and we implement this within a thermodynamic cycle workflow, which then means I have a starting well type in the unbound and the bound state.
33:36
I have the mutant in the unbound and bound state.
33:38
It's hard in silico to calculate energies for binding and unbinding, but since we have a thermodynamical cycle, it's much easier to calculate these two arms.
33:49
So impacts of mutation in the unbound state, impacts of the mutation in the bound state, which then means that I can actually rearrange the equation and get the binding free energy value.
34:01
How does it perform?
34:03
This is just an example of three papers that shows you the correlation between the experimental data and the FEP prediction in case of predicting impacts of mutation for stability and for affinity.
34:16
I think for Affinity, we have an extra paper now like that was published last year where FEP was validated on 12 different systems, including six antibody systems and one TCR systems and other systems as well.
34:32
OK, now we switch back here and I'll show you maybe if I go here and I would like to perform FEP calculations how this can be done.
34:46
Let's just quickly go here.
34:48
This is the FEP protein mutation.
34:49
It looks quite similar to the MM-GBSA panel, while actually it looks similar, but like the type of calculations happening in the background is different.
35:01
And again you just I have the 1912 structure selected here.
35:09
I load it now I have a list of different of list of all the amino acids that I can actually test impacts of mutations on.
35:18
It's a binding calculation because as I mentioned, you could calculate predict impacts of mutations for stability or affinity.
35:27
And now I can, I need to define what is binding to what.
35:30
So in this case, it should be like the heavy on light chains binding to the other chains.
35:35
And now I have a list of the heavy on light chains residues and I can just again go to heavy chain 30 and 31.
35:42
And again, just maybe this is out of select the mutations to be calculated.
35:50
OK, Or as I showed you pass it from the residue scanning one and it's going to show up once you start when you then you need to start the calculation and then the results will look like this.
36:06
This is how the FEP result would look like.
36:09
We don't have time to go through all the details, but I can show you like now you have a list of each mutation that has been calculated and then you have the predicted impact of the affinity on the affinity with uncertainty value on it.
36:27
And then you also have the representations.
36:31
For example, this is a map that shows you like, OK, this is the wild type and it was mutated to this one and this one was mutated to this one.
36:38
Like you know, all the mutations that have been or the perturbations that have been calculated.
36:44
I showed you in the benchmark 3 papers.
36:48
If you have the experimental data, for example, if you're running a benchmark or like an internal project, you want to validate your calculation.
36:55
You can actually add your experimental data here and then you can plot such a graph.
37:02
Important thing here is also you in addition to this, you get more information about the impact of mutation in the impact of mutation on like on interactions and so on.
37:14
So you get the energy values.
37:15
But if I click here, for example, I get this is the mutations that was introduced.
37:22
This is the impact of this mutation on the interactions of the original amino acid with its neighbouring amino acids and the target amino acid with its neighbouring amino acid.
37:33
You get more details about the convergence and like basically a lot of details about the statistics of how this calculation was done gives you more confident in your results.
37:47
OK, so this is how you run FEP.
38:11
So I want to cover two more things just to so those are new features that we basically added on to cover more and more or like address more like address some of the challenges that we saw while calculating impacts of mutations using FEP.
38:31
The first one is one of the actually important things is can I actually engineer a variant that is pH sensitive, which means that I can I actually predict impacts that actually would allow me to tell, ah, I can have this mutant and this mutant will be binding at a certain pH value and unbinding in the and not binding that much in the in a different pH value.
38:57
This is, if you think about it, this is actually possible having the infrastructure we discussed.
39:02
Because the difference is I need a pH value that I will predict the impact of mutation on.
39:08
And then I need to just see if I transform a protonated to deprotonated or the other way around, how this is going to impact the binding in the same way.
39:18
What we saw from our benchmark is now this is what we discussed.
39:22
This is the similar dynamical cycle that we discussed.
39:25
But what if the protonation of the titratable side chain in the unbound state is different from when it's binding because simply the side chain, let's say a histidine, when it's unbound, it's solvent exposed.
39:43
But once it's bound, it's now hidden at the interface, which might cause some PKA shifts, which then would result that basically the protonation, the preferred protonation state would be different.
39:57
And this is now done by just simply.
40:00
We added an extra step of calculation that basically then calculate the impact of this shift in the and, consider this in the result of the calculated delta delta G.
40:16
Just two examples of if you run what we call naive FEP, which when you didn't, we didn't look at this.
40:23
And if you run the corrected FEP and how the result or the value of delta delta G will be compared to the experimental value.
40:35
Second example here that this is like the staphylococcal like an antibody binding to Staphylococcus enterotoxins and the idea there was actually that a histidine was identified that can be APH sensor in there that you we predicted that at a pH 7 I yeah 7.4 versus a pH 6 this the protonation state of this will modulate the binding to its entity.
41:07
This is one part quickly.
41:10
The second part I want to cover today is I mentioned that MM-GBSA is qualitative but fast, so allows you to maybe use it to just exclude the non-option amino acid mutations FEP accurate to the experimental level, but it needs time and it needs compute resources.
41:35
So there was always the question from scientists like, OK, do we have somewhere in between some tool that would allow you to maybe more accurately assist the impact of mutations, but it doesn't require that much compute resources.
41:55
And now we have what we call the residue scanning FEP.
41:58
So residue scanning FEP, it's in the middle, it's MD based.
42:02
So all concerns about like if there is adaptation needed for the introduced mutation, you need, you have discovered here because it's running with MD simulation, but it's also not gradually transforming your protein, your amino acid from 2.
42:23
But it's actually having a thermodynamic probability of transitioning from 1 mutation from 1 amino acid to the other, which can be calculated much faster.
42:35
But at the same time, since it's probability, it is actually much more accurate than MM-GBSA.
42:44
The way this goes, this is how FEP works.
42:49
And residue scanning would be instead of like several steps, you have just probability being calculated to transition from the wild type to the mutant.
42:58
And since actually this is done in one step, sort of it's continuous, then you can run this not only on one side on several sites as well, which means in an experimental language, it means that you now can also look into multiple mutation.
43:16
I can actually look into the probability of having several mutations and several bindings, sorry, several positions of the protein, how good it is.
43:30
We basically looked at the 12 systems in the paper we published last year, and you could see looking at the last right hand side columns how the RMSE values are quite similar to FEP.
43:49
Another representation of this is like on the blue, you have the blue dots, the FEP, the red dots.
43:57
This is the new method that we that I'm actually introducing now and again.
44:03
The idea there is basically then that this will be much more accurate than MM-GBSA, which then means that what you can do, you have a starting, let's say you I have several mutations, hundreds of mutations, hundreds of ideas.
44:17
I can use MM-GBSA to quickly exclude the non like not optional like the bad ones and then go to a more accurate, not very computationally expensive one to even further filter those and then go only with maybe the much more confident ones to FEP.
44:50
So yeah, for protein, FEP as some has said much more accurate, but it's much more computationally costly.
44:57
And so if we're going directly to FEP, then we're spending a lot of our compute resource in this top right part of the graph, which is mutations which are bad for affinity and predicted to be bad for affinity.
45:09
So this is where methods like MM-GBSA and FEP residue scanning can actually help us enrich and get the, you know, more important residues higher in our list so that when we do go to the more accurate methods that we, you know, that we're doing it in an efficient manner.
45:26
So just to contextualise this, so as part of a collaboration, we had an experimental partner and they made, did a saturation mutagenesis on 39 interfacial sites for an antibody system.
45:41
And they, so they made tested this and they had all the ELISA measurements we did in parallel with our technologies in this workflow.
45:50
So let's imagine you take those 39 sites, you mutate it to all amino acids except cysteine.
45:56
OK, Now, just as a benchmark, you would need 8 96 well plates to cover that.
46:02
Now, something that you don't know a priori, which, but from my experience, you see it's around the hit rate, if you will, good mutations is around 5%.
46:09
So what it means is in any single one of those wells, you have a 5% chance, in fact a 4.7% chance of finding an interesting mutation.
46:16
And this is something that's -0.5 kilocal per moles around 2 to 4 fold binding or better in any one of those wells.
46:24
The benefit being that all of your good mutations are within that panel.
46:28
But then if we take our FEP residue scanning, in this case, we actually exclude proline mutations to and from.
46:34
This is something that we're developing at the moment to improve that.
46:37
So the number goes down, but this takes around 5 days on around 40 GPUs, which you can access with, you know, cloud providers.
46:47
We then take those top, you know, we rank those, we take the top 250 from the FEP rest just getting into FEP to quantify them.
46:54
And this takes a bit longer, around 2 1/2 weeks on 40 GPUs.
46:58
But this in terms of compute cost is competitive with experiment and I'll show you that in the next slide.
47:02
So let's say that we took not 8 96 well plates, but just 1 96 well plates.
47:07
Well, it turns out that of the 33 variance active variance, 21 are in there, which means that instead of a 4.7% chance of finding a good mutation at any one of those wells, you now have a 22% chance.
47:20
So you've increased the enrichment fivefold, and it means that also 64% of all the possible hits are within the panel and again within a few weeks with eight times fewer experiments.
47:31
And when we compare this workflow that I've just described on the right hand side with what you typically get from sort of walk through mutagenesis with the CRO.
47:39
And again, we're just basing this on quotes that we've had in our research in the area that this can take, you know, for such scale around six months to do.
47:48
Again, they need to make, test, analyse and you know, purify all of these proteins.
47:52
And we're taking this as 100% of the cost to do the workflow that we just mentioned.
47:57
It can take, you know, around 1/3 of the time and at 20% of the cost.
48:01
And that's for all the technology and the cost to run it.
48:05
So hopefully that gives you sort of a bit of context where, you know, we can use this computational workflow to augment your protein design process and save you know, money because it's much cheaper time because you can get the results faster and you have better outcomes because you can also couple these with some of the developability, you know, calculations that you've done before.
48:25
This might be great for Infinity and stability, but it's terrible for, you know, for aggregation, for example.
48:32
And so with that, I just want to round off for this session and the first part here and offer time for some questions.
[PART 2]
59:23
OK, I think we'll get started for the second half of the workshop.
59:28
Thank you again very much for sticking around.
59:32
For those that you have that have and anyone that's joined, then welcome.
59:37
So in the second part of this, we're really going to be focusing on, you know, bringing our data, you know, experimental data and all of our other types of data together to make better decisions, also to analyse it.
59:51
So our enterprise informatics platform is what we call it is called LiveDesign.
59:57
And so this has been around for I think around 10 years at least and is, you know, basically used in all majors, big pharma clients for small molecules.
1:00:09
So this is where it has been, you know, ideating and tracking molecules and so on.
1:00:13
And we've recently extended this functionality to biologics and so.
1:00:19
The idea is that this is a web-based platform and this allows you to be, you know, working collaboratively with people.
1:00:28
So if you especially if you're across multiple sites and you're working on the same project, you can be reading the same thing, having all the same data available to everyone.
1:00:37
You can have it closed.
1:00:38
You can also have it closed in specific projects.
1:00:40
We have that, you know, control as well.
1:00:45
And The thing is you have the key benefit here is that you know, I think anyone that's worked in this field, maybe I'll take a show of hands who has sent a protein sequence to their colleagues as an Excel file.
1:01:03
Does anyone enjoy, you know, actually tracking that?
1:01:07
And again, that's in a separate file.
1:01:10
And this is where the platform comes in, having all of that data in one place so that you have your sequence data, your in silico data, your experimental data, your structures, all in one place and all connected so that you can understand that in a very fluent way so that you can make the decisions.
1:01:27
It's a agnostic platform.
1:01:30
Now, of course, we've mentioned the technology with that we use, but if you have, you know, your favourite thing for, I don't know, HPLC or something, these things can be connected.
1:01:39
And so it's really, it's a platform to integrate all of the data.
1:01:43
And that's the philosophy that we've taken.
1:01:45
And so again, just to show you here in the what it's like in small molecules conceptualising how we take, you know, this DMTA cycle type thing and that you can actually have all of that data there, make predictions, docking with just all of this stuff and have that with your wet lab assets along with your structure.
1:02:04
And again, just to mention that we have this platform layer, we have the interface that's then you know, we can have our shorting our tools below that and certain views and visualise things.
1:02:12
I think it's very powerful, especially for visualising the data and having that at hand.
1:02:18
But then of course having all of the API that you can put into other tools that you use frequently.
1:02:26
And of course we do this in the extension to biologics because the DMTA cycle with small molecules is this fairly simple one that you see on the left hand side of the screen for biologics.
1:02:35
It can be quite different and not a lot of things are working in tandem that you need to go back and forth.
1:02:40
And so tracking that can be quite difficult.
1:02:43
And of course, if you have this in, you know, Excel files and spreadsheets and so on, it's not really, you know, it can become a bit intractable.
1:02:51
And so this is where we see the benefit of having this integrated platform for having everything in one place.
1:02:58
And yeah, just to we'll talk mostly on proteins, but we do support nucleic acid bases.
1:03:04
So DNA, RNA also nonstandard amino acids, all of these formats as well.
1:03:09
And we have quite a lot of features, but we'll talk about really in the context of some examples today.
1:03:16
And so I think with that I'm going to hand over to my colleague Ilaria.
1:03:19
And yeah, again, any questions, put your hand up and then we'll try and get round to you.
1:03:25
Ilaria.
1:03:28
Thank you, Dan for the introduction to LiveDesign.
1:03:30
And I'm just going to briefly going to walk you through a sort of, it's a like a demo or a click along situation.
1:03:40
So sorry not click along, but a demo situation where I will do the clicking and you can see all of the things that we can do in LiveDesign in terms of data that we're going to look inside of the platform.
1:03:51
I'm going to be talking mostly about antibodies, and we are going to take again the same MEDI structures that Assam presented in the first part.
1:04:02
And what we are going, what I will show you basically is how we can generate a lot of mutants to these antibody structures and then effectively analyse them in with interactive plots and other type of filters to certain parameters that we want to look at.
1:04:20
So the way that after we analyse our antibody structures, we can also triage them efficiently in order to identify the best possible candidates that will then be brought further for, you know, therapeutic development.
1:04:37
So the way that we are going to generate these mutants is by the same residue scanning with MM-GBSA that as I mentioned previously, we're going to look at the results from the residue scanning and also other aspects such as developability parameters.
1:04:53
We also touched upon these previously in the first part, we're going to talk about the aggregation score and also the TAP properties to see how these antibody mutants can efficiently become have the possibility to become therapeutics.
1:05:09
And we're also going to briefly touch upon the FEP results.
1:05:13
So the same in the method to calculate stability and affinity of your structures.
1:05:21
And as I said, the same, our starting point for this short demo is the same data that was prevented previously.
1:05:28
So we already talked abundantly about the MEDI 1912 and 578 structures.
1:05:35
Just as a little refresher, these have up to 12 mutations, most of them are localised in the heavy chains and three of these mutations were shown to give a higher aggregation score in a specific site on the MEDI 1912 structure.
1:05:53
And what we are going to do then in LiveDesign is to use these structure and make a selection on which type of candidate we would like to bring further for mutant generation and then triaging all of the mutants generated.
1:06:09
So we're going to have again a look at all of these parameters and these will help us to make decisions.
1:06:18
And we already talked about, you know, the back mutation from the three residues that had been identified and how, yeah, how basically the same mutations can efficiently be analysed with the AG score method that is the same one that we have implemented in LiveDesign.
1:06:35
And how useful it is to have these in silico predictions because then you can relate them obviously to your experimental information, right?
1:06:45
So let's go to our LiveDesign platform.
1:06:50
So hopefully the text is not too small for all of you.
1:06:55
So this is how the platform looks like.
1:06:59
We usually can select a specific project that we work want to work on.
1:07:03
In this case, we prepared this folder called MEDI antibodies.
1:07:07
So this is all the people that are connected to this project or have access to this project will be able to use the live reports in that live inside of this folder.
1:07:16
So a live report, I can show you an example maybe to first keep things organised.
1:07:22
I'm going to create my own folder on the LiveDesign so that I know where all of my data is going to be stored.
1:07:31
So that's MEDI sorry tutorial with my initials just to keep things clear.
1:07:39
And this is what's one of our live reports looks like, because this is the sort of master copy, I'm going to actually duplicate this.
1:07:49
And again, my initials, I'm going to save it in the folder that I just created so I can start touching the data on this live report and forget about the previous one that I closed.
1:08:01
So here are the two antibody variants loaded on our page.
1:08:08
And as you can see, we have the experimental information.
1:08:11
So we have a KD value and also an experimental delta G value which is loaded in the database and is brought back into the live report page.
1:08:22
We also have the experimental data indicating formation of higher order sort of assemblies and also the aggregation score which is connected to the formation of these higher order assemblies.
1:08:36
So we saw previously in the plot that the two values were correlating.
1:08:41
We could also have a look directly at the 3D structure of these antibodies by hovering over this image.
1:08:46
But more sort of organised way to have all of this data into one place is by setting a certain form.
1:08:55
So these forms are layouts that can be saved as templates, and they are very useful when you want to introduce new data into your live report so that it will be added into this to this starting template.
1:09:08
So now I have my entities or my antibodies here on the left and now I have activated them or selected them.
1:09:18
So now we can sort of start analysing what some aspects.
1:09:24
So at the bottom of this form you have the sequence viewer.
1:09:29
We have two panels.
1:09:30
So in one case you could be looking at the heavy chain of the antibodies and we would need to align the two sequences.
1:09:36
In the other panel, you could be looking at the light chain of the antibodies and again align the two sequences.
1:09:42
You can give them a colour.
1:09:43
For example, we're going to choose very standard hydrophobicity scheme.
1:09:50
And now you would be able to scroll through the sequences and quickly identify which residues mutate between the two antibodies.
1:09:59
So for example, here for the heavy chain we have, we can see that we have a 90, roughly 93% percentage of sequence, sorry, of homology between the two sequences because that's where most of those 12 mutations were localised.
1:10:14
Whereas for the light chain, there's 100% of homology.
1:10:20
Now in terms of 3D structure, we can activate this visualisation that we have loaded for the structure if the connection wants to work, yes.
1:10:33
So here are the two antibody structures represented as with a surface and we can see the same patch given by the aggregation score algorithm which is just colour coded.
1:10:47
So for the first, the 578 antibody, it's quite light because as we mentioned before, it doesn't show very high aggregation propensity in this region.
1:10:57
Whereas if we switch on the other one, you can see the same paths that we were looking at previously with very high scores.
1:11:05
So with all of this information in hand, you would now be able to sort of make a much more informed decision on what type of candidate you want to bring forward for development.
1:11:18
And the type of sort of development that we're going to do, as I mentioned is ready to scanning mutagenesis.
1:11:25
So for the sake of time now I'm just going to select the first candidate, and we are going to proceed with this through the mutagenesis and then the final triaging and looking at different aspects of these mutants.
1:11:41
So to do this, we can directly copy this, the information related to this structure into a new live report where we perform the residue scanning analysis.
1:11:52
So I'm going to call this with the antibody name and then residue scanning and again my initials.
1:12:01
We can also give it a template which determines the type of columns that you're going to visualise.
1:12:07
You can save these templates again.
1:12:08
These are very useful when you have new data incoming.
1:12:11
And we save this in the folder that I generated, right?
1:12:19
So here is our structure.
1:12:21
And as I mentioned this form was set up to perform in very easy manner.
1:12:27
Now for demonstration the residue scanning analysis.
1:12:30
So we have columns that indicate which positions we want to scan to which mutation we want to mutate to and then this will be the running.
1:12:39
So the button where to click run so that the calculation is performed.
1:12:42
The residues will be brought back into the following columns.
1:12:45
So they have already been prepared here.
1:12:49
We still have the 3D view as well as the aggregation score information.
1:12:55
And we also included the top analysis parameters.
1:13:00
So it's the same ones that Dan mentioned before about these sort of five key parameters that define different aspects of your antibody structure.
1:13:10
The way we treat them here in the live report is sort of like what we call a multi parameter optimization issue.
1:13:18
So we have, as you see, these 5 scores that then become combined into one score that you have at the end and you can recognise it with the sort of summation sums symbol.
1:13:31
And then the score is has will have a value in scaled between zero and one.
1:13:37
So obviously 0 means that all of the parameters are not performing very well.
1:13:43
Whereas if you're closer to 1, most of your parameters are getting closer to what you would want for an antibody therapeutic, right?
1:13:52
So to perform the residue scanning, we can come to the scanning positions and we can click exactly on the positions that we want.
1:14:04
I'd like to click more and more on.
1:14:05
So I have, I think I remember the ones that were causing belong to the aggregation paths should be these three positions.
1:14:15
Again, just for the just because this is an example, we're going to select those three positions and mutate them, for example, to an alanine if we want to perform the alanine scanning.
1:14:28
And then we're going to run the calculation and basically just wait for the results to be brought back.
1:14:37
So the calculation has actually already been run previously.
1:14:40
We are just bringing back the results into the live report because there wouldn't be time now to run a new calculation.
1:14:48
So you see now these three mutants and the columns give you very valuable information of for example, which what was the original residue to which residue it has been mutated to the position that it occupies in the sequence.
1:15:02
And then as I meant as I mentioned the results coming from the residue scan calculation.
1:15:09
So these are the same ones that as some was talking about in the first part of the presentation.
1:15:15
So we have our delta affinity and delta stability.
1:15:18
They indicate a change in stability and affinity in the mutant structure with respect to the wild type being a sort of global energy of the system.
1:15:29
We want to have negative values.
1:15:31
So for example, this already is an antibody candidate that improves its stability, whereas in terms of affinity it hasn't really changed much from the wild type.
1:15:42
It actually got a little bit worse.
1:15:43
And in terms of aggregation, it kept very similar aggregation propensity values.
1:15:49
And also the top analysis seems to be quite OK.
1:15:53
So we don't just need to limit ourselves to 3 antibody variants.
1:15:59
The power of LiveDesign is in fact that you can triage a whole bunch of data.
1:16:05
So you can generate much a much larger data set.
1:16:09
And in fact, I'm just going to select all of the residue positions for the heavy chains because we have recalculated the values for this one.
1:16:18
And we can also, for example, show all of the possible mutations.
1:16:25
So you could be doing this as you know as a fresh calculation in this case it’s just going to again bring back the values from the database.
1:16:39
And you can see that now we have much more antibodies.
1:16:42
We have actually generated 550 antibody variants and now you start to see even more variability in terms of the results.
1:16:51
So you have change the different types of changes, for example, even to worse delta affinity or delta stability scores.
1:17:00
And then at the same time, you could be having a look at your MPO columns.
1:17:05
So the top parameters here.
1:17:08
So instead of just having a list of results, we could again be analysing these structures with a basically interactive plots that also these can be saved in these forms.
1:17:23
So in these sort of template layouts that then can be used in other when adding other data.
1:17:31
So we have set up a plot here on the left that shows the delta affinity coming from the residue scanning method on the X axis and the delta stability on the Y axis.
1:17:43
As I mentioned, we want ideally to have negative values for both of these results.
1:17:49
So what we can do is to select if I manage if I to select the lower left quadrant in this plot and although the compounds selected here are going to be brought back in the plot on the right, which is the aggregation score on the Y axis.
1:18:06
So we know from you know our information on the starting candidate which had an aggregation score of around 65 or 66 that we just want to improve the, we don't want to get worse in aggregation propensity, right?
1:18:22
So we are going to select all of the structures that have an aggregation score of 67 or even better, so even lower.
1:18:36
And this basically selects the structures that are then brought back at the bottom list.
1:18:42
So now we have filtered down from the 550 candidates to roughly a hundred 116 structures to even be a little bit more clear or clean.
1:18:54
On the filters, we actually have a filtering tab.
1:18:58
So you could use here the, again, the residue scan results.
1:19:03
So we're going to apply 2 filters coming from that.
1:19:07
And that's the stability.
1:19:09
As I said, we want 0 for both of them, right?
1:19:19
So you see that the selection becomes even more clean.
1:19:22
We can also add the aggregation score filtering.
1:19:31
Right.
1:19:32
And now we really got down to structures that we are really interested in which are around 70 antibody structures and to have further information.
1:19:44
We also sorry, let me toggle off the filters actually.
1:19:49
So these can be very easily toggled on and off.
1:19:51
So now we go back to the starting situation.
1:19:54
If you want to add more structures to your filtered list, you could just, you know, control click and select additional structures from the plot.
1:20:02
So the plots on the list are always interactive and this really helps you to make your decisions so.
1:20:11
Finally, another form that I wanted to show you is this form to have a better look at the mutants that you have created.
1:20:20
So for each position on the sequence, we can then have a look at the type of mutant that it was created.
1:20:27
And the this heat map or these, you know, yeah, scatter plot is coloured by the delta affinity coming from the residue scan and sized by the delta stability.
1:20:38
So again, this is another source of information.
1:20:42
You want basically green or you know, green, yellowish dots that are also, I think it should be small and you can see that the larger red ones have been excluded because those are the ones that we didn't want.
1:20:56
Similarly, when you are looking at the aggregation score, again for each sequence position, you want the smaller dots that are coloured in green because they have the lower aggregation score values.
1:21:10
So again, if you're interested in a specific position and you want to on your sequence and you want to include a certain amino acid mutation, you could just control click and add this to the list.
1:21:20
And the final list that we have the at the bottom is the one that is going to be passed on to for further analysis.
1:21:29
The type of analysis that we can do further that we would propose to do further at this point is basically the protein FEP method.
1:21:37
And so this checks in a much more accurate way for stability and affinity of your antibody complex.
1:21:47
So you can see that here we have a column that we called pass to FEP.
1:21:52
You can add this column as a sort of free form or you know, yeah, it's a sort of, it's a special type of column that you can do several things with it.
1:22:00
You can either add a check or a cross whether you like the structure or not.
1:22:06
So this is like a yes or a no option, or you can, it also can be a column where to add your own comments and it's just a free form column that can have different purposes.
1:22:17
In this case, we checked with a yes, all the structures that we would like to bring forward in the FEP method, which is going to be, you know, it's computationally more intensive.
1:22:27
So you really don't want to pass all of the structures that you had at the beginning.
1:22:31
You only want to select the really good promising candidates that you have analysed and you have triaged after all of these analysis and plots that we have shown.
1:22:43
So to show you just the results of the FEP, we can open again an existing live report where the results have been from the FEP have been imported starting from the candidate, the same candidate that we chose at the beginning, so the MEDI 578.
1:23:05
So I'm just going to open this one for just demonstration purposes.
1:23:12
So these would be the same antibodies that we selected from the previous analysis, and they have been brought back because you can import the structure based on the yes or no pass that we assigned before.
1:23:26
And in the first columns we have the results coming from the FEP.
1:23:31
You can see again, so the affinity free in the free energy for affinity and the free energy difference for the stability of the structure.
1:23:41
Again, because this method includes dynamics and much more sort of accurate calculations, the results can be a little bit different from the residue scanning, which is a slightly more approximative method.
1:23:53
But it was useful at the beginning to triage all of the candidates.
1:23:57
And at this point you would be following the FEP information to understand if your antibody mutant has improved in terms of affinity and stability.
1:24:10
So if it is actually a better lead to design your therapeutic structure and yeah, you still have all of the previous information.
1:24:22
So you have the aggregation scores and still the TAP parameters.
1:24:27
And hopefully this is a way that allows you to really know what is going on with your data in a much more sort of confined manner in the sense you have everything in one place and you can make informed decisions based on this.
1:24:43
And I think that's the end of the demonstration that we had for today.
1:24:48
So if you have questions, we're very happy to answer them.
1:24:52
Could you go back to the MEDI 578 RES scan?
1:24:56
Yeah, that one now again, just to show how.
1:25:02
Yeah.
1:25:03
So if you scroll to line 4 for one second, yeah, maybe it's not there.
1:25:14
Give me one second because I want to make a comment on one of the properties here.
1:25:19
But yeah, so I'm just adding this.
1:25:24
So if you Scroll down, I added a comment to the mutation, which is line three.
1:25:30
Yeah.
1:25:30
So for example, I've just now added this comment said we don't like cysteine and so on.
1:25:35
If you're collaborating, you could be on a Zoom call and you can also make decisions and say, oh, no, we made that before something happened, or this is difficult.
1:25:43
We don't, you know, you can do this live.
1:25:46
And that's the really powerful part of it.
1:25:48
So again, yeah, this is where it's worked really well for chemistry teams and small molecules as a design.
1:25:55
But here's how we can also bring into biologics and add up things in a live manner.
1:26:00
Exactly.
1:26:01
You can also check if, like other people, in this case Dan, are acting on your live report.
1:26:05
Because up here it would flag up this little sort of plus one, or it could be +2 or whatever people that are working on the same live report.
1:26:15
So it's highly collaborative.
1:26:28
OK.
1:26:28
Are there any questions on the platform itself or if you want to continue talking about the methods that are being run in the background, which are most of them would be the same, you know, core suite methods that we have available?
1:26:54
OK.
1:26:54
If not, I guess then I don't know if you have some closing comments.
1:26:59
Yeah.
1:27:00
So if you go back to the slides, yeah, then yeah.
1:27:03
So just to say, I mean, of course there's some questions that come here and hopefully then this is just one example in one workflow, but it's really flexible.
1:27:15
And so hopefully what you can see is that bringing together the sequence of structure, all of that data in one place can be really powerful.
1:27:25
And one thing that we didn't maybe highlight too much here, which is quite important is that now we've glossed over this, but it's aware that it's an antibody Fe.
1:27:38
It's also where what is the heavy chain was the light chain.
1:27:41
And so each of these will have their own identifiers.
1:27:43
So like if you're having a common light chain, it will have that common light chain there.
1:27:47
It's all tracked.
1:27:49
Now you will be given, if you're ideating, you'll be giving a new generic identifier.
1:27:54
But if you have your own corporate identifiers, etcetera, those are also tracked.
1:27:59
You will get multiple identifiers.
1:28:01
So it's quite powerful.
1:28:03
It deals and I'm going to go off pieced here a touch.
1:28:08
So this was I think Project 1 yes yeah.
1:28:17
So let me just see if I can find there was format.
1:28:22
So I can't remember off the top of my head.
1:28:28
I need to find the specific multiple modalities with hierarchy.
1:28:35
Thank you, Joe.
1:28:37
So this is just an example.
1:28:38
It uses HELM and this is especially useful for things like peptides.
1:28:44
So here we have a branch peptide and it knows what is the branch, what is the trunk, because it has that topological knowledge that is built into HELM.
1:28:55
We're actually planning also a builder where you can actually make these connections, you know, for not just backbone, but also, you know, cysteine bridges and other things.
1:29:03
That's also in the works.
1:29:05
Oh, an error that's my fault.
1:29:08
And it again works with DNA and RNA.
1:29:10
So it knows the different sides here.
1:29:13
I think this has caused kicked off the network.
1:29:16
Yep.
1:29:16
So it's kicked off the network, but that's my fault.
1:29:21
But hopefully what you can see from this is that it can deal with also bispecifics.
1:29:24
It can it knows topologically, OK, this is a heavy chain, this is a light chain or this is one side of the antibody that's made of these.
1:29:31
And so you can really go down and break down your system into different components.
1:29:37
And so it's aware of that.
1:29:38
And of course, tracks that all through.
1:29:40
So, yeah.
1:29:42
And this finally did.
1:29:43
I managed to get the image up.
1:29:45
So you are here.
1:29:47
And if you make your way through, this is where you find our booth #48 And yeah, that's what it looks like.
1:29:52
Very subtle, as you can see.
1:29:54
But please do come and talk to us about, you know, any ideas, comments or anything.
1:29:59
We'd love to hear from you.
1:30:00
So again, I want to thank you all for your attention and thank you to my colleagues for presenting today.
1:30:06
And of course, happy to take any questions now or later.
1:30:09
Thank you.
