Computational Methods in Structure Based Biologics Engineering

0:29 

Our discussion group leaders for today are Leonardo De Maria, who is Principal Scientist at AstraZeneca and Origene Nyanguile, who is Professor at the University of Applied Sciences in Western Switzerland. 

 
1:05 
This is the first time Origene and myself are involved in one of these things. 

 
1:12 
So you are our guinea pigs for that. 

2:08 
And I think what, what we agreed to do was just to give you the framework for the discussion and then moderate the discussion. 

 
2:21 
So it's up to you how things are going to go. 

 
2:24 
So I think the intro is going to be quick by me. 

 
2:29 
And then there are a few starting topics which is around machine learning, protein structural prediction of individual proteins or protein-protein interactions. 

 
2:42 
And then also how in general can you use these tools to optimise biologics. 

 
2:50 
And also maybe we could, depending on what you are interested in and how the discussion goes, it could also be about the tools themselves, if they are viable and if they are usable by people or not. 

 
3:04 
But first I just wanted to make sure that we are all on the same page. 

 
3:16 
And of course biologics, if I say biologics, everyone thinks about antibodies, right, which are the most used biologics, but they are by no means the only biologics out there. 

 
3:27 
And this is by no means an exhaustive list of biologics or alternative scaffolds, right, that are out there. 

 
3:38 
Some are more famous than others. 

 
3:40 
DARPins, probably them very well. 

 
3:44 
They are coming from the lab on Professor Andreas Plückthun in Switzerland. 

 
3:51 
You might have encountered anticalins because Pieris is a company that is building a portfolio around anticalins. 

 
3:59 
And the whole point here is that as in the antibodies, you have the variable regions here, the CDRs, that you can use to search for binders. 

 
4:16 
In these other scaffolds what the positions that you see here in red are also in a similar way randomised to find binders. 

 
4:25 
And there are pros and cons on both type of scaffolds and also even if you stick to antibodies. 

 
4:33 
So the complexity of the formats of antibodies is very large. 

 
4:38 
The classical mono violent antibody is just one of the many possibilities that you have. 

 
4:51 
But this is just to put the framework on what biologics could mean. 

 
4:58 
And then I go to an extreme vision, and the extreme vision which maybe reflects at some point sometimes the expectation from top management is in my eyes the following. 

 
5:17 
So you have an antigen where the sequence of the antigen. 

 
5:23 
You might be very specific about an epitope. 

 
5:26 
So you might be very specific about a region of the sequence of this antigen. 

 
5:32 
You want to find binders because that way they will modulate the biology you are you're interested in. 

 
5:40 
Then you have a series of requirements, which is how strong the binding affinity you need. 

 
5:48 
You might know that already beforehand if you have done your PK PD analysis or not. 

 
5:55 
You may want to do antibodies, but you may want to go to any of the other biologics that I just told you. 

 
6:02 
Let's stick to antibodies for a time. 

 
6:05 
You might decide already what is the administration route typically sub Q for antibodies these days. 

 
6:11 
But I'm in respiratory and in immunology we tend to for respiratory also think about inhalation. 

 
6:20 
Of course, yes, you may have ideas about your PK profile. 

 
6:24 
Then once you start to think about on the line, you may want also depending on all these things to have requirements on how good your expression yield needs to be at the end of the day, once you have landed into a candidate. 

 
6:38 
And then also some requirements about the viscosity, so on and so forth. 

 
6:43 
So you may have a series of requirements, these sequence of your antigen, maybe with some details. 

 
6:51 
And then there comes machine learning. 

 
6:53 
There is this big algorithm and after some computational time, you end up having a sequence or a set of sequence limited that you will try in the lab and that they will fulfil these things. 

 
7:08 
If you take the vision and put it in an extreme way and we could say that’s top management vision and maybe some of you are already in top management in your company. 

 
7:19 
So and then what could be a little bit more realistic? 

 
7:25 
What you see is that instead of the big robot in there, you have a number of algorithms that can address a few of the things that are here. 

 
7:33 
And you will see in one of the examples I have later on how this could be done. 

 
7:39 
So you can have some algorithms that are good at classifying binders versus nonbinders. 

 
7:45 
You may have some algorithms that are good at predicting solubility. 

 
7:49 
You might have some other algorithms that are good at predicting immunogenicity. 

 
7:53 
And all of a sudden you get into a combination of these algorithms that will again reduce the number of sequences that you can try. 

 
8:02 
So this is this is the extreme, the super extreme vision. 

 
8:08 
OK. 

 
8:09 
And then let's go a little bit more into a few steps backwards. 

 
8:15 
And then if we think about protein structure prediction, which is very relevant either for the antigen so if you want to know what is the structure of the antigen that your biologics needs to bind, or if you want to know the structure of your biologics itself. 

 
8:30 
And this is definitely, I think everyone in the audience is aware of the huge changes that happen in the field, mostly due to the AlphaFold in two algorithms from DeepMind, which essentially has to a very big extent cracked the problem, which is given a sequence, you have the algorithm, and you have some structures that are predicted. 

 
9:01 
And actually these are real predictions that I tried, not from these, but from other proteins using AlphaFold. 

 
9:07 
And this is now available to everyone, AlphaFold. 

 
9:11 
So everyone even in industry can get the programme and if you have GPU's available, you can run and then you can start to play with these predictions. 

 
9:20 
And there the computations are a little bit… it takes a few hours for one prediction. 

 
9:27 
So it's not really super-fast, and part of it is because they use a lot of evolutionary information. 

 
9:38 
So they use sequence alignments to derive some structural constraints coevolution. 

 
9:45 
But what is happening now? 

 
9:47 
And this is something that Professor AlQuraishi from Columbia University has published lately, but also the new algorithms from Facebook from Meta are doing is this multiple sequence alignment independent methods using protein language models, which is essentially what if you have the text completion algorithm in your phone, when you start typing, based on the initial words, it will by probabilistic guessing, so to speak, see what is the next word? 

 
10:23 
It's the same with the amino acids. 

 
10:25 
And these methods are faster and can be as accurate as the more computationally intense methods. 

 
10:32 
And there is also very interesting initiative, which is called OpenFold, which is a group of industries essentially getting together where again, Professor AlQuraishi is the scientific leader on that initiative. 

 
10:50 
And then to develop methods for protein structural prediction that is essentially outside of DeepMind. 

 
10:59 
And how is it when you want to go to protein-protein interactions? There are already some methods available there. 

 
11:08 
So AlphaFold has something called AlphaFold multimer, which definitely works much better than some of the standard protein-protein interaction methods that you see here. 

 
11:19 
And you also see here in the diagram, some very good predictions from the AlphaFold multimer. 

 
11:28 
There are some also predictions for antibody antigens with some variable accuracy. 

 
11:38 
But this is if you ask if you talk to the guys from DeepMind, this is where they want to go next. 

 
11:45 
So this is where they are going to focus their efforts. 

 
11:47 
So you will count on seeing more and more of these methods where you can get a very good prediction of protein-protein interactions with computational methods, essentially starting from the sequence of protein A and protein B. And which goes to the next part, which is once you have derived these things. 

 
12:08 
So once you essentially start from maybe a repertoire of sequence of antibodies and maybe your antigen and you can start to do computational work in order to address what we computational chemists call the pose of the complex. 

 
12:23 
So how the two things are going to interact together. 

 
12:26 
And then you can even start to think about and do some in silico affinity maturation or affinity predictions and try to use some methods to predict the effect of mutations on the binding affinities. 

 
12:39 
And there are some methods available already there. 

 
12:41 
They are very computationally intensive. 

 
12:44 
They are called free energy perturbation methods. 

 
12:47 
And you could use them with some with some success to predict the changes in ΔΔGs upon a mutation on a protein-protein complex. 

 
12:59 
And then the final slide I have as an introduction is this one here from a recent paper from a group in Switzerland, which is actually very interesting because besides the specifics, it highlights more a realistic vision of how things might work. 

 
13:20 
And, how things might work is that you have a specific protein-protein interaction that you want to modulate with antibodies because of different reasons, you run an initial experimental campaign. 

 
13:35 
In this case, of initially they did a size saturation library and based on the size saturation library and some binding data, they went into another more tailor-made library that they screen. 

 
13:55 
And then you use the experimental data that you generate yourself in house to build a machine learning model for your own need. 

 
14:06 
In this case, what they build was a classifier of binding of this antibody to HER2, essentially. 

 
14:14 
And once you have these, you can have your library screen in silico to select binders versus nonbinders. 

 
14:23 
And once you have your enriched library and the library was… you can play with the 10 positions at the same time. 

 
14:33 
Here it was experimentally one position at a time in saturation. 

 
14:37 
And then once you use you have your library in silico that you can classify for binders versus binders with a model that you have built in house. 

 
14:50 
And then you add other filters that are developability filters and maybe immunogenicity filters. 

 
14:56 
You can end up with sequences that you will try in the lab. 

 
14:59 
So this is something that is actually more realistic and happening. 

 
15:05 
And it's a balance between the extreme vision I showed you where computers do all and you just do a little bit of experimental work. 

 
15:13 
And here would you have a nice interplay with between experimental work and computational work. 

 
15:33 
And then I think now we can start the discussion. 

 
16:19 
Yes, I'm Phil Noble. 

 
16:21 
I'm a senior group leader at UCB Pharma based in Slough. 

 
16:27 
I have a small group in antibody discovery, but on the wet lab side of things. 

 
16:31 
So I don't know too much about structural biology and computational methods other than the stuff that we've been doing here at UCB. 

 
16:41 
But, I'm obviously acutely aware of the way that the field of antibody discovery is moving. 

 
16:47 
And as you mentioned people are expecting de novo antibody design at some point. 

 
16:52 
And as I want to make sure that we stay at the cutting edge of the field because we have a pretty good wet lab capability in this area. 

 
17:03 
So I was wondering what people's thoughts were on people's claims in the literature and companies claims about de novo antibody design, because it's very difficult because the field is moving so fast. 

 
17:18 
It's quite a lot of hype around some companies’ massive deals with big pharma and some biotechs in this space. 

 
17:26 
And I was wondering if anyone had any hands on experience or being able to validate those types of claims and whether we are really actually near a near that Holy Grail of de novo discovery or more like you were suggesting Leonardo, where we need to generate a lot of wet lab data to then try and drive those models. 

 
17:59 
So my name is Andreas Evers. 

 
18:01 
I've been working a lot of years at Sanofi doing computer aided drug design on small molecules and then on peptides. 

 
18:09 
And now I'm moved to Merck, Germany, which is not MSD and working on biologics. 

 
18:17 
So referring to what Leonardo has presented, I made also good experience, or we made good experience with what you have shown on the last slides from Derek Mason with having a lot of data and giving these active sequence data to let's say generative models and let that suggest new sequences. 

 
18:38 
Now for the docking and the virtual screening. 

 
18:42 
When I started to work on small molecules for my PhD thesis, I thought… it was in 2000 or so. 

 
18:50 
I thought in five years, we will have solved the docking problem and the scoring function, and we can dock all small molecules into all proteins and simply identify binders. 

 
19:01 
In publications it worked, in prospective studies, when I did a virtual screen for small molecules in my hands, I had hit rates of between 2 and 5%, meaning you test 1000 sequences, you get 50 actives and activity means also 200 nanomolar or so. 

 
19:19 
And I know Leonardo, there was also at that time a presentation from Stefan Schmitt, who had been working at AstraZeneca on modelling. 

 
19:27 
He had exactly the same experiences. 

 
19:31 
And when I talked to my small molecule modelling colleagues here in my company, my impression is that docking and scoring has not yet significantly improved. 

 
19:41 
So maybe we would now get hit rate of between 5 to 10%. 

 
19:45 
Now I think the small molecule docking problem is much less complex than the large molecule docking problem. 

 
19:52 
I just did a few experiments in silico experiments where I tried to redock crystal structures. 

 
19:58 
This worked, but when docking a crystal structure onto the same target but taken from another PDB, I was not able to reproduce the binding mode because of some small side chain rearrangements. 

 
20:11 
I think this is something maybe that AlphaFold could use quite soon in the future. 

 
20:16 
But I think in the small molecules we have not yet solved the scoring problem. 

 
20:21 
And this is something that I think will be quite critical. 

 
20:25 
Let's assume we can dock one million or billions of sequences or antibodies or let's say therapeutics onto a target. 

 
20:35 
What would be the hit rate? 

 
20:36 
And in a realistic scenario, if I go to my colleagues and I say, let's produce 100 of those antibodies, we expect a hit rate of 2% and 5%. 

 
20:47 
And now let's say from 100 sequences, two will show a binding affinity of 200 nanomolar. 

 
20:55 
Where to where to continue here? 

 
20:57 
So this is at the moment where I'm a little bit sceptical. 

 
21:01 
I know that quite recently there has been two or three evaluations on AlphaFold docking. 

 
21:07 
And I know also that the study from the Charlotte Deane group about ZDock, where the conclusion was docking, you can enrich known active binders by docking into crystal structures. 

 
21:18 
But in both studies I think the conclusion was that there was room for improvement when docking on model structures. 

 
21:25 
And then once we have solved that one, I'm really a little bit concerned about just solving the scoring or the affinity prediction problem on a huge data set, not just docking 10 sequences, but maybe to dock millions of sequences or thousands. 

 
21:44 
So this is my personal opinion. 

 
21:45 
I don't have any concrete experience, but I always say to my colleagues, let's look what the world is doing. 

 
21:53 
And if there are first success stories about prospective applications, then I will also start to work on such approaches. 

 
22:10 
My name is Jeffrey Luo, and I work for J&J in Spring House, USA and I work mostly on proteins, antibody structures, and antibody engineering projects. 

 
22:25 
I used to have a crystallography group. 

 
22:27 
So I’m familiar with the antibody structures and how they interact with, with antigens and so on. 

 
22:34 
But I was also curious actually, let's say I have a question for Andreas that you just mentioned about a few in silico experiments that you've done where you try to do docking of antibody antigen. 

 
22:49 
If you split the complex and you do docking, yes, you can easily get back to the crystal complex. 

 
22:58 
But if you take one of the components or both components from different crystal structures, it's very hard to generate the correct pose and find it in the large number of models that you can generate. 

 
23:14 
So one question I had was in your experience, or if anyone has experience, when you make models using either AlphaFold or with the other ones for known crystal structures with complexes. 

 
23:32 
And then you do whatever docking that you try and of course, you can use the crystal complex to find if there are any poses or any models that are similar to the crystal structure. 

 
23:43 
In those cases, do you actually have these models in the ensemble that you generate that are ranked high that would resemble in a reasonable manner the crystal structure. 

 
23:59 
The reason that you can't find, of course, they have relatively poor clashes and so on. 

 
24:03 
So the scores may not be very good, but if they're not even in the vicinity of the correct complex, then you'll never get there. 

 
24:11 
So in your experience, you see these things that actually, yes, they are there, it's just I can't find them. 

 
24:21 
That's a separate question then if you never get them close enough, then you will never find them. 

 
24:29 
Yeah, I can directly briefly answer on that. 

 
24:31 
I just tried on one case, redocking the crystal structures and then generating models. 

 
24:37 
Crystal structures were consistently found on rank one. 

 
24:40 
The models were found on rank three or rank 10. 

 
24:43 
So based on that, I said OK, for virtual screening at the moment, with the workflow that I used, it might not work, but I did homology modelling or antibody modelling with MO. 

 
24:55 
I just used one docking tool. 

 
24:57 
Then I said for the tools that we have internally available, it would not be feasible. Maybe with AlphaFold, but still, still let's say in my hands, I would not dare to make a virtual screening based on homology model structures. 

 
25:13 
And I think this was also conclusion from the AlphaFold virtual screening or redocking evaluation because I think they said in the paper that the CDR-3 modelling accuracy is also with AlphaFold not... 

 
25:28 
I don't want to say anything wrong, but I think they stated there was still room for improvement. 

 
25:35 
I hope this answers your question from what I said in between the lines. 

 
25:41 
No, no, thank you. 

 
25:42 
Thank you so much for that. 

 
25:45 
The other question I have is, you also did a lot of small molecule work. 

 
25:53 
So if we take models, I guess I haven't really done much exploring myself, but I was told that the models for target proteins let's say, even antibodies, they tend to be out of AlphaFold 2 for example, are quite accurate. 

 
26:12 
My question to this audience is, if you compare them to crystal structures and then you can measure the RMSDs and whatever. 

 
26:28 
Do you think that the models coming out of AlphaFold 2 or whatever is good enough even for small molecule drug discovery work? 

 
26:42 
I my own very crude to look at it is maybe yeah. 

 
26:49 
So you can see it's probably the fold. 

 
26:52 
But there are many inaccuracies in the details where, for example, the active sites and things like that, they may not be very useful for in silico docking small molecule screening, for example, virtual screening. 

 
27:05 
So what is your take on this? 

 
27:09 
Yeah, I think I will add to what Andreas said before that the devil is in the details. 

 
27:15 
Because it's more about what do you do with these models. 

 
27:20 
So for a small molecule, would you take one of these models and then run, as Andreas was hinting, a very expensive computationally virtual screening of millions of compounds on that specific AlphaFold generated structure? 

 
27:34 
Or would you take this antigen Fab structure that AlphaFold multimer has predicted and then do computationally expensive affinity maturation and then take those suggestions seriously. 

 
27:47 
And for small molecules the devil is in the details. 

 
27:51 
It might be that the side chains that AlphaFold for a reason gives you are those that are more abundant in what AlphaFold has seen, but probably is not the one for that binding pose. 

 
28:06 
So it will take you out off the right path. 

 
28:12 
And the same could be for that. 

 
28:15 
And again, you could use AlphaFold to generate the structures for molecular preparation in X-ray crystallography. 

 
28:24 
So they are actually very good structures, right, but good enough to do drug discovery? 

 
28:31 
I think it's a question mark. 

 
28:33 
I think there is no clear cut answer to handle with care. 

 
28:47 
So I'm Andrew Buchanan, a long-time colleague of Leonardo. 

 
28:51 
I do biologics discovery for Cambridge Antibody, Medimmune, and now AstraZeneca for over 20 years and since 2018 I've been evaluating stroke exploring this whole computational structure of biology, machine learning, application for making antibodies into medicines. 

 
29:14 
And I've worked for a long time with Charlotte Deane's group and other structural biology groups. 

 
29:21 
I used to lead a structural biology group. 

 
29:22 
I don't at the minute. 

 
29:23 
Thankfully this frees up a headspace to do all this stuff. 

 
29:26 
But I think there are a few really good questions that you're entering in one is how good is the docking? 

 
29:34 
My view is it's not good enough. 

 
29:36 
AlphaFold for me is a toy. 

 
29:38 
It's no good until it can do PPI's. 

 
29:41 
I post that all the time. 

 
29:42 
Whenever DeepMind posts anything I say great. 

 
29:45 
It has great applications if you want to understand an unknown gene structure function idea. 

 
29:52 
But if you want to play in protein-protein interactions, it cannot do it. 

 
29:56 
Well, even AFM can't do it. 

 
30:02 
It will get there one day, but it can't do it because there's nothing to learn. 

 
30:04 
All AlphaFold does is at a grand scale, spots sequence motifs and the folds that go with them. 

 
30:11 
That's all it does. 

 
30:13 
It does it at grand scale and it is a great application when you want to work in an unknown space. 

 
30:20 
But if you want to do protein-protein interactions, you can't do it with AlphaFold 2 or AlphaFold M in my view at the minute. 

 
30:30 
So when it comes to docking, the tools aren't good enough when it comes to predicting affinity. 

 
30:36 
Actually, some of the tools are brilliant. 

 
30:39 
If you have a crystal structure and you use open source FEP, you can make serious headway really quickly. 

 
30:52 
We haven't published it yet. 

 
30:53 
I'm trying to encourage my colleagues to publish it, but it's brilliant. 

 
30:58 
One drawback though, Andrew with that comment, is that open source, I mean FEP in general. 

 
31:01 
It is open source or commercially provided. 

 
31:04 
It's very computationally intensive. 

 
31:06 
So you definitely you cannot screen libraries. 

 
31:14 
Yeah, but you can make more than site direct like talk to Juan Carlos's team, they do brilliant stuff. 

 
31:21 
So this is a bit of an internal chat. 

 
31:23 
Yeah. 

 
31:24 
You cannot screen for 100 variants. 

 
31:28 
You could screen 100 variants really quickly. 

 
31:30 
In fact, you can make libraries of libraries. 

 
31:33 
That all depends on your computational power guys. 

 
31:36 
Obviously at AZ we can burn the diesel. 

 
31:41 
Anyway, that's FEP, FEP will make great headway. 

 
31:45 
If you have really high quality structures, then it comes to machine learning, the only way you can get it to work well, you should look at publications in 2021, 2022. 

 
31:59 
There was another really nice paper doing machine learning for antibodies, essentially about antibody LO, you can do it. 

 
32:08 
You can do it for antibodies, you can do for peptides. 

 
32:10 
My group will publish. 

 
32:13 
Let's hope we get something in Nature Chemistry in the next quarter, which basically demonstrates again if you've got data, you can machine learn anything that is related to sequence and function. 

 
32:28 
You need data. 

 
32:29 
You did really well curated data. 

 
32:31 
How many of us have that in our industries and groups? 

 
32:33 
Not many of us, not even AZ does it well. 

 
32:37 
And if you go to any of these future lab informatics ELN type conferences, everybody says the same. 

 
32:44 
None of us do data well enough because it was never meant to be ready for machine learning. 

 
32:48 
It was meant to help make decisions on drug projects. 

 
32:52 
But you can do it. 

 
32:53 
And machine learning will let you LO faster and explore space that you went in. 

 
33:01 
Last comment then is where do you go with structured, guided, rational design? By which I mean define a structural epitope and design computationally an antibody-like molecule, or a peptide that will bind that. 

 
33:18 
Two groups have published it. 

 
33:20 
The best one is Sormanni in the Cambridge Department of Chemistry. 

 
33:24 
He made antibody-like molecules. 

 
33:27 
The other group that does it is Baker and there's another Swiss group that do it, but they use rigid bodies and they're molecules. 

 
33:33 
I bet you are as immunogenic as you would never take that to a patient. 

 
33:36 
It's very, very elegant when you work in the rigid body space, but the rigid body space is not the flexible antibody space. 

 
33:44 
That's all I have to say at the moment. 

33:50 
I think that's good. 

 
33:51 
Those are the points that people want to hear. 

 
33:57 
And this is from real world experience. 

 
34:12 
And in the pharma, and even though you haven't published, hopefully you guys will do that soon. 

 
34:28 
I have a question for Andrew. 

 
34:30 
I fully agree. 

 
34:34 
You mentioned that with open source FEP, you had good results. 

 
34:38 
And is this for different antibodies or is this more if you have one crystal structure and you design in mutants to do an affinity maturation? 

 
34:47 
So this is something we are currently exploring actually not with open source tools. 

 
34:51 
But do you mean you can compare the binding affinity, let's say of different antibodies binding to different epitopes, or rather if you have one epitope to find mutations that for activity maturation? 

 
35:06 
So this is for the latter. 

 
35:08 
So basically this is when you have one high quality, let's say to less than 2.5 angstroms resolution, the paratope and the epitope and you can engineer around it really, really well. 

 
35:22 
I will add, Andreas, that this is a very similar case. 

 
35:27 
We do a lot of that also for small molecules. 

 
35:30 
So you could use FEP as well for small molecules where the change that you do is not an amino acid mutation, but it's just a functional group and things like that. 

 
35:52 
So then over there sometimes it works very, very well. 

 
35:58 
So you could definitely get a very good correlation with experimental data. 

 
36:01 
And sometimes it does not. 

 
36:04 
And there are some clear culprits about when it doesn't work and when it does. 

 
36:11 
And our experience as well with the biologics is the same. 

 
36:16 
Sometimes, it works well and sometimes it does not. 

 
36:21 
I'm not agreeing, Leonardo. 

 
36:25 
Leonardo and I sit in different but related groups. 

 
36:31 
And my firm belief as I experience this space more is that biologics are, let's say they're alive, they're not self-replicating. 

 
36:42 
But BLOSUM is a real phenomenon. 

 
36:44 
If what BLOSUM is, amino acids know how to relate to each other. 

 
36:47 
That is a real language in biology. 

 
36:50 
Chemistry doesn't have an equivalent. 

 
36:54 
Is that true? 

 
36:56 
Yes, it is true. 

 
36:57 
You need to know BLOSUM is. 

 
36:58 
If you don't know what BLOSUM is, go and read a textbook. 

 
37:03 
The no, my question was, are you sure these small molecules, they don’t have different functional groups, and they don't talk to each other? 

 
37:12 
Oh, no, that's not what I mean. 

 
37:14 
I mean a language, like we're speaking English, and we know what the next word's going to be. 

 
37:18 
And how they relate to each other. 

 
37:20 
Amino acids know how to relate to each other. 

 
37:23 
They follow rules. 

 
37:24 
We call it BLOSUM and other things. 

 
37:26 
Anyway, this is this is more a philosophical point, but I believe that's why machine learning for biologics will work in a way that is much more impactful and fast than it will ever for chemistry because biologics have been in this space for maybe four or five years and they're making headway in a way that chemistry hasn't in 15 years. 

 
37:47 
That's very pretty of provocative, but I'll just put it out there. 

 
37:51 
So can I ask my follow up question to that when I have quite a bit of experience with FVP. 

 
38:03 
Of course it's based on like a proprietary algorithm that we pay a lot to run right. 

 
38:08 
So is the implementation that is proprietary the implementation? 

 
38:12 
Yes, the implementation. 

 
38:23 
You start from a very good structure. 

 
38:24 
And my question is if you if you didn't have the co-structure with the 2.5 Angstrom resolution would you be able to do similar things with models when the models are less accurate? 

 
38:41 
No, my experience is don't bother. 

 
38:44 
We've tried, and my experience is don't bother. 

 
38:47 
OK, good. 

 
38:52 
You can see one of the publications where we did it with models. 

 
38:56 
It was awful, but it got published. Don't know why. I'll try and post it in here if I can find the thing. 

 
39:03 
Oh, that would be great. 

 
39:05 
And just to comment on the FVP, I realised I've been working small molecules, peptides. 

 
39:10 
Then I moved to biologics. And in small molecules, if you have a small molecule screening hit and you have your crystallisation platform established, you might have a crystal structure quite soon after that. 

 
39:23 
But imagine you do an animal immunisation, you have some hits, then you have to produce that one as a Fab and then you have to co-crystallise. 

 
39:33 
And usually the timelines we obtain the crystal structure is usually after the time we have available for sequence optimization. 

 
39:42 
So this is just another strategic aspect. 

 
39:44 
When I joined the company, I thought, great, we can leverage X-ray crystallography. 

 
39:50 
Now after two and a half years, I say, yeah, let's try to make the crystal structure as a backup. 

 
39:55 
But if everything works within the timelines that we have, we will not be able to make use of the crystal structure for affinity optimization, of course for epitope definition, that's fine. 

 
40:07 
But I also have to admit we are working with external collaborators and maybe this is the reason why the process of crystal structure generation takes so long. 

 
40:16 
I think that's a really, really good point and it is a bone of contention again between chemistry and biologics generally in the industry. Chemistry are well resourced for structure, we aren't: biologics aren't. 

 
40:30 
Just generally in the industry because chemistry is used to driving SAR. 

 
40:36 
But given the breakthroughs as FEP, I am pushing my company to give biologics resource and they will give us it if they believe we'll do the same thing as chemistry does, which is basically use it. 

 
40:50 
Whereas you're right, the way we normally use it for biologics is for pretty pictures and epitope definition at the end. 

 
40:56 
Is this everybody's experience? 

 
40:58 
So is this your experience that you don't get enough structural biology support if you're doing biologics? 

 
41:04 
So is everyone on the same boat? 

 
41:07 
I agree it takes a lot of time, but I do think we get a lot of support in in the biologic space. 

 
41:16 
It was quite impactful on one of our drugs actually. 

 
41:19 
We're able to engineer in other isoforms specificity for our anti R17 antibodies. 

 
41:26 
But that's an old story, Phil, that's from years ago and you never published it. 

 
41:33 
It only sits in your patent. 

 
41:34 
It doesn't sit in the public domain, but it was really good. 

 
41:39 
In that case, we did require the structure to help achieve that. 

 
41:42 
And I guess the key is the complex of complexity of the biology. 

 
41:48 
If you just want an IL-17 binder that found wherever, then you're never going to catch up and need to be able to use that crystal structure. 

 
41:56 
But if you if you have the time to do that then it makes it a lot easier. But yeah, we are trying to push for more structural biology earlier on projects so that we can actually impact with some of these things because we can make antibodies better and better therapeutics. 

 
42:13 
So is cryo-EM helping? 

 
42:16 
So can you use cryo-EM to get faster complexes of Fabs and antigens? 

 
42:24 
Yes, definitely. 

 
42:25 
Two against one. Andrew, you are in a minority here. 

 
42:32 
I have Chris Phillips. 

 
42:33 
Tell me what he thinks. 

 
42:34 
I'm trusting Chris. 

 
42:35 
I don't know what we published. 

 
42:37 
I'm not going to say any more than it does help. 

 
42:40 
Hi, this is Medha Tomlinson from Takeda. 

 
42:44 
I think it's certainly evolving. 

 
42:46 
And as Andrew mentioned the impact is really very high. 

 
42:51 
But I do agree that for biologics, they need to drive the design. 

 
42:57 
That's when the impact is higher. 

 
42:59 
And to that point, we are looking at cryo-EM and though not published, I think it's still emerging here in Takeda. 

 
43:07 
But I think because we're a fairly new group in biologics, but I do think in the future, this and the AI and other aspects there's always the conventional ways of getting antibodies, but to enable that differentiation, and that IP space, we need to proactively use these tools to really make sure we drive that function. 

 
43:31 
And when you're able to do that upfront as opposed to post getting a lead and working on it, I think that's what's going to help. 

 
43:39 
And prior to Takeda, I was with AbbVie for many years and certainly they have an awesome structural group that has been doing a lot of that. 

 
43:48 
But there's as this field emerges and the new machine learning takes some of this aspect of designing and how we can put these leads, then I think it's going to be a very exciting field. 

 
44:09 
I certainly I led a structural biology group trying to do as many structures as we can as fast as we can to support the ongoing discovery and the optimization processes. 

 
44:23 
And I think for a while we were really successful until of course things changed organizationally. 

 
44:31 
But what I want to say is that I still want to come back to the point the Andrew mentioned about the using open FVP to do serious computations based on high resolution crystal structures for affinity optimization or whatever the other properties that you do. From my experience, going from doing this structure to support engineering versus actually being responsible for the engineering itself. 

 
45:04 
And I have to have my people go into the lab to make those variants and to show that they actually do work or improve upon the previous ones. 

 
45:15 
So my interactions with the many people in the experimental side is, it's a lot faster to use methods that are, for example, like a phage display, yeast display, or even mammalian display to optimise these antibodies much faster than your computational work can do. 

 
45:38 
So what is the real value other than having a pretty beautiful story and a pretty picture starting from a crystal structure? 

 
45:53 
So you have fired all your comchems essentially? 

 
45:56 
That's for this reason that my people, my structure people were fired was because this reason. 

 
46:03 
So our optimization people can do very well without the structure. 

 
46:10 
Why do we need the structure? 

 
46:12 
Well, epitope definition for sure. 

 
46:14 
That's why we still have them. 

 
46:16 
For that you need one structure. 

 
46:19 
So at the end of the day, for the present application. 

 
46:22 
So yes, it's great to be able to calculate the properties, your affinity and then design variants, but those can be very quickly explored experimentally in phage and yeast. 

 
46:36 
So how, when you are faced with those arguments, I'm not saying we don't do structures, don't do computations. 

 
46:45 
I like them. 

 
46:45 
And that was many years of my work. 

 
46:50 
So because in companies or in anywhere, you have different sides. 

 
46:58 
How do you respond or deal with that argument? 

 
47:04 
So Jinquan, the way I think I'd answer your question is we discussed the benefit of structure, now with the tools you have available, but there are lots of things you can do without structure. 

 
47:17 
You can epitope map. 

 
47:19 
You don't need structure to do most of what we do because it's all driven by biology. 

 
47:24 
You don't need to be a chemist. 

 
47:25 
You need to not think like a chemist to make this world work for biologics. 

 
47:34 
And how do you address the developability things down the line. 

 
47:39 
So that by empirical levels, you just look at it. 

 
47:44 
I'm just asking Jeffrey. 

 
47:50 
Certainly when you make an antibody, you can just do all kinds of measurements and figure out what properties it has, if it's viscous, it aggregates… 

 
48:01 
And then typically that's from the from CDRs and you think in your experimental design. 

 
48:11 
And then you select for the ones that are less so by whatever methods. 

 
48:17 
It takes trial and error. And so, the difficulty that I have with my colleagues previously and now and then, I'm pretty sure in the future, is that there will always be these different sides. 

 
48:34 
And of course in addition to science, there's also politics. 

 
48:39 
How do you convince those people who are sceptical about this kind of work? 

 
48:45 
I mean, regarding the structural biology, what is easiest to convince is success stories and when I joined the company, I even took responsibility to start crystallisation as early as possible. 

 
48:59 
Let's say we have a set of ten hits and people will try to profile them down to two or three hits. 

 
49:05 
I already asked to produce those hits as a Fab, which might take at least six weeks and then realistic scenarios, it might take three months. 

 
49:15 
Then you send them over for crystallisation. 

 
49:17 
You will have the crystal structure earliest, six months later and then in between, I will be asked for those three years to start sequence optimization. 

 
49:27 
People will not ask to wait for the crystal structure because we have time pressure. 

 
49:32 
So for me, it's always a backup for the second design cycle. 

 
49:36 
And just three weeks ago, there was the lucky situation that the sequences have not been potent enough. 

 
49:42 
And in that week we got a crystal structure, and I started FEP already. 

 
49:47 
This might be a success case, but this might happen in one of three projects or one of five projects. 

 
49:53 
So in between, if we get a sequence, let's say from the llama or mouse, we do the standard humanization, hoping that we will not lose too much affinity. 

 
50:03 
And now referring to your other question, for that purpose, I'm also using in silico developability tools, but I consider them to have a similar value as Lipinski's rule of five. 

 
50:15 
Let's say we can eliminate the rubbish, and we can make sure that we don't put the totally poor guys into in vitro developability assays. 

 
50:23 
But yeah, we are heavily using in silico developability, but not to replace these assays, let's say to filter down from 100 sequences to 10 sequences. Or to say, look, from this one, they are very similar. 

 
50:37 
Please take the one that has a lower predicted hydrophobicity and so on. 

 
50:42 
But this is something that we try to establish being fully aware that viscosity is difficult to predict in particular in specific formulations. 

 
50:52 
But I'm pretty sure that with the tools we can eliminate the poor guys. 

 
50:58 
I think we now certainly strongly established that good structural support. 

 
51:06 
Yeah, actual crystal structure or high resolution structures are really important. 

 
51:12 
So I mean of course, the purpose of this discussion is, is how AlphaFold and the information from there. 

 
51:24 
Can we do most of it with that information rather than getting crystal structures, EM structures, they take time and effort. 

 
51:37 
So, how close are we to being able to use that information in most of the work? 

 
51:44 
Not for affinity maturation, for example, but like predicting solubility and patches that we can engineer and focus on to engineer how close are we using these models? 

 
52:01 
For solubility predictions, Jeffrey, we published some time ago when I was at Novo a paper with Michele Vendruscolo on CamSol. 

 
52:10 
And on that paper we compared a number of computational methods to predict solubility of antibodies versus a large number of experimental methods on a large set of variants of a specific antibody. 

 
52:26 
And there we could show that CamSol works better than all the computational methods and as good as many of the experimental methods. 

 
52:38 
And CamSol, that's not the structural information. 

 
52:42 
Actually you can run CamSol based on only the sequence of your antibody. 

 
52:48 
So you could add a structural spice into it, but in principle, the predictions we had that were just sequence based. 

 
52:57 
So as a long term biologics engineer, CamSol is interesting and there are a few other fairly good developability predictors computationally. 

 
53:08 
But we're the wrong people to answer this question. 

 
53:11 
This is really a question for our colleagues who work in which you might call CMC or biopharmaceutical development BPD, the teams who have to deliver formulation at 100 to 200 mg/ml. 

 
53:21 
That's a big deal. 

 
53:22 
That's a formulation question and it's developability question and viscosity question. 

 
53:29 
They're the only teams who have the data to do that work. 

 
53:32 
So everything we do in developability is, as I think Andreas said, we're just weeding out the rubbish. 

 
53:38 
That's a different question to can you turn this thing into a medicine at a high concentration liquid formulation. 

 
53:46 
That's not our problem. 

 
53:49 
It is a real world problem, but we're not going to answer it. 

 
53:52 
Our CMC/BPD colleagues have to answer it because it's a whole other area of expertise. 

 
53:58 
CamSol’s fine, but it's not it's not good enough. 

 
54:02 
But you will see the BPD groups from groups like J&J… 

 
54:12 
A few of the big pharma BPD groups have published some really nice papers this last three years.  

54:21 

I’m assuming kings, King’s brief from Sanofi about colloidal specifically. 

 
54:24 
Yeah, that was a nice paper. 

 
54:27 
This is for me also the most predictive one. 

 
54:30 
They said there was one descriptor that captures 80% of developability properties. 

 
54:36 
But different companies and different people have different developability. 

 
54:41 
For one, it's also immunogenicity. 

 
54:43 
For the other one, it's PK. 

 
54:46 
Yeah, they're all different aspects. 

 
54:48 
I would think of developability just within the manufacture piece because immunogenicity is a whole other field, as is PK. 

 
54:56 
They're whole different assays and they're outside of BPDs remit, but they are. 

 
55:00 
Can you turn this concept into a medicine? This drug into a medicine? 

 
55:07 
I work in those spaces, but I haven't got a good answer for immunogenicity or for PK yet because to do that, you need to do what's called bedside, back to bench. 

 
55:19 
All the time people like us are pushing data and molecules into the clinic from the bench to the bedside. 

 
55:24 
Hardly ever does our businesses enable bedside to bench. 

 
55:29 
I am actually pushing AZ really hard to get that data. 

 
55:33 
But it's ethically hard and scientifically hard and data process hard because nothing is set up to give you the data down. 

 
55:41 
It's only there to give data up, in terms of to regulators and super senior management. 

 
55:46 
If you want actual data, I don't know how to get it yet. 

 
56:00 
I wanted to answer the question from Jinquan. 

 
56:09 
The scope of AlphaFold to predict developability properties. 

 
56:13 
My personal belief is those predictions are not accurate enough that they depend so much on the quality of the model. 

 
56:20 
If I may, I use Xscore, but I could use any other tool for hydrophobicity prediction. 

 
56:25 
My belief is if I use the model from MO or from immune builder or AlphaFold or from a crystal structure, it doesn't make a big difference on the predicted hydrophobicity. 

 
56:36 
And therefore I also agree with Andrew. 

 
56:40 
I would express it in a more diplomatic way, but in principle, AlphaFold predictions are a toy for me. 

 
56:46 
You can look at the structure, but just looking at the structure doesn't bring me a new drug. 

 
56:51 
And also regarding would it be interesting for, in principle, the docking. 

 
56:58 
And I had the chance once to talk to John Jumper from DeepMind. 

 
57:02 
And I asked him what do you think about docking and predicting of oligomers? 

 
57:07 
And he said AlphaFold was not designed for that. 

 
57:11 
Yeah, I fully agree. 

 
57:12 
So he didn't blame me for that question. 

 
57:14 
And I don't blame AlphaFold for not being suited for virtual screening because it was not made for that. 

 
57:21 
That that's great. 

 
57:22 
To me, that is similar to my own experience. 

 
57:29 
I should say we have a few minutes left. 

 
57:32 
Can I ask one more question? 

 
57:33 
I've been the one that's asking a lot of questions. 

 
57:37 
So hopefully these questions are relevant to others as well. 

 
57:39 
So, one thing is that when we think about developability, we typically look at surface features and hydrophobicity charge patches and so on. 

 
57:50 
But for proteins that fold from sequence to this three-dimensional structure, the folding itself is a kinetic process and the antibody, or your biologic solution is also a kinetic dynamic beast. 

 
58:05 
And we know this folding even though AlphaFold and the other algorithms can predict the final overall 3D structure pretty well, but what happens from sequence to the end or anything in between that is not very clear from the protein folding process. 

 
58:29 
And that has a lot to do with the developability in some cases. 

 
58:34 
That's my personal view. 

 
58:35 
So what do you think? 

 
58:37 
Yeah, so what do you guys think, has the AlphaFold solved the protein folding problems, so to speak? 

 
58:46 
No, it didn't aim at that. 

 
58:53 
That was not the goal. The goal of the programme was not to predict the way you get to that folded structure. 

 
59:02 
But that's the way Leonardo, the external world reads it like it did do it even though it didn't do it and it didn't and intend to do it. 

 
59:08 
It's completely overspun. 

 
59:10 
Why are you so pro AlphaFold? 

 
59:13 
What I was saying is that that's not what the programme was meant to be, which is I think what Andrew was saying, and people were saying, yeah, it solved the folding problem. 

 
59:26 
It didn't solve the folding problem, it just solved the protein structure prediction problem, which is from sequence to structure. 

 
59:33 
But what's in the middle is an algorithm. 

 
59:35 
It's not the folding of the sequence to that specific protein structure. 

 
59:54 
I mean if nobody else has something, Jinquan, you mentioned the folding aspect. 

 
1:00:01 
I'm not sure. 

 
1:00:01 
So I personally don't know whether folding is so important for developability. 

 
1:00:07 
Of course we shouldn't have antibodies that are unstable at 40°, but I often hear people saying that folding stability correlates with long term stability. And this is actually to my knowledge not the case because I made long term stability predictions and whether your molecule unfolds at 60 or 70° does not correlate with long term stability in a pen device. 

 
1:00:33 
I can imagine if you go for the virus inactivation step, then it's critical to have some folding stability. 

 
1:00:39 
But I personally ignore a little bit the folding stability. 

 
1:00:44 
But it would be very interesting to hear from others how they… 

 
1:00:51 
I just put a paper on the chat. 

 
1:00:52 
It's an old paper, very old, from 2010. 

 
1:00:54 
It's from, from a great physical chemistry in Spain, Professor Jose Manuel Sanchez-Ruiz. 

 
1:01:01 
And there he goes in length kinetic versus thermodynamic stability. 

 
1:01:09 
And what we can measure in the lab often is thermodynamic stability, but what happens in the tube is kinetic stability, which is harder to measure. 

 
1:01:17 
So that's a very oldie but goody paper. 

 
1:01:21 
But Leonardo, from a biologics perspective, good folding does relate to everything that's good about an antibody. 

 
1:01:29 
If there's anything that's not about the rate of folding, the TM isn't terribly predictive. 

 
1:01:36 
But yeah, if it doesn't fold well, and you're given the business like a $10 million problem to solve. 

 
1:01:44 
Why would you ever do it? 

 
1:01:47 
Or in another way. 

 
1:01:48 
I asked my CMC colleague, should I care if one antibody has TM of 59° and the other one of 65°? 

 
1:01:58 
And he said ‘meh.’ 

 
1:02:08 
I thank you so much for the discussion. This is really helpful to me. 

 
1:02:09 
And I can say there are cases where that TM difference is important. 

 
1:02:15 
Yeah, of course, 10° difference. 

 
1:02:18 
And if we talk about 45 compared to 60°, I fully, fully agree. 

 
1:02:30 
Thank you so much, everybody.