Founder Stories: Ben Holland, Computational Antibody Discovery

So let me introduce Ben Holland, the co-founder of Antiverse, Oxford-trained and developing novel AI machine learning methods to bring therapeutics. Yeah, so please hear from Ben.

Thank you. Thanks, Konrad. Can everyone hear me okay?

Yeah, there is some I think that there is some kind of noise. I don't know if it's just me.

Okay.

I think that's good now.

Okay, great. So, as Conrad said, I'm here from Antiverse, CTO, co-founder. We were founded in 2017. I'm here to talk about the title of my talk, Machine Learning Based Design of Antibodies Against Difficult Targets, which is just one of the types places where machine learning can contribute significantly to computational antibody discovery pipeline. Antiverse was founded to change the course of patients' lives by engineering new biologics, and we also founded on the basis that machine learning could be a big part of that. So we've been machine learning from the beginning, we're still here, and I'm glad to talk to you today. The structure of the talk: I'm going to talk about why do computational antibody discovery and what parts of that we work on, what it can offer when it works. And then in the middle section, I'm going to talk about kind of 4 angles looking at the problem, 4 lenses through the problem, if you like, from a computational point of view. I'll then go and talk about current challenges in the field, where Antiverse comes in, how we've gone about things and our progress, and then where I see the future challenges. So why would we do computational antibody discovery? In general, when you've got a system that works, basically works a lot of the time, if you're thinking of adding something new in, you're looking for benefits over that approach. And we think there are places where computational antibody discovery can provide significant benefits. One thing we've noticed, as I think all the speakers have noticed, is that when you combine experimental work and computational work, you get the best of both worlds. I don't see computational work being kind of evolving independently of experimental work or coming to replace it, except in certain narrow cases. Current methods are diverse and sophisticated, they work really well, and so we need to think carefully about what benefits we can do and where we can fit together and where the synergies are. Traditionally, people will talk about the kind of 3 types of improvements to make to a system. You can improve quality, you can improve speed, and you can improve cost. Obviously, speed can be highly valuable. If an antibody comes to market and brings in, say, a billion dollars of sales per year, 6 months' saving can be very valuable, even if that saving is 5 to 10 years in the future. Quality is a multi-dimensional thing. We've spoken about a lot of things about that, particularly developability is key. I think that is such a diverse issue that each individual customer or consumer has their own idea of what the priorities are for quality, but we've noted a few of them today. In principle, we can reduce costs by using machine learning, but I think most of the time quality and speed will trump pure kind of cost reductions for most therapeutic purposes at least. One other dimension that isn't kind of covered by those is new capabilities. So immunization and phage technology, but display technologies generally, they cover a lot of targets. And if you've got a recombinant target, so a soluble target, you can usually, almost always, for a good lab, you can find antibodies against it. But there are certain cases where there are traditionally considered difficult targets, and an example, classic example, a transmembrane proteins, and, well, there's a whole class. But the idea behind Antiverse is that computational methods can help and can collaborate with those methods, and we can open up, uh, in one— in the words of, um, Sandeep, I think he put it as increasing the antibody-druggable space, which is great, great way of doing it. So, um, there are 400 more GPCRs and ion channels associated with many diseases. There are only 2 FDA-approved antibodies against GPCRs specifically. So these are— I think most people would agree these are a difficult drug class, and it's difficult to discover antibodies against them for lots of reasons. 2 of them I pick out on this slide are cells. Because they're transmembrane proteins, you have to put them on a, on a cell, a living cell. And with a few exceptions, it's extremely difficult to get them to be soluble enough to take other approaches. When you put things on a cell, there are lots of other proteins and receptors on the cell membrane that kind of get in the way. We call that the forest problem, trying to find the receptors you want amongst this forest of other receptors. And antibody libraries with poor specificity. That's not to denigrate other antibody libraries, but when you have an antibody library with maybe 10 to the 8, 10 to the 12 antibodies in it, that's a tiny subsampling of antibody space, and that makes it difficult as well. So Antiverse is focusing on hit generation for difficult targets, and with a specific focus at the moment on GPCRs. The reason for that is we want to get therapeutics to people, not particularly get them faster or cheaper to people, but we want to make drugs that couldn't otherwise exist for practical reasons. So when we say we want to make antibodies against difficult targets, that brings us to kind of terminology. We call it de novo discovery, and that means starting from nothing. De novo, from new. And that is a particularly tricky case because by definition, by my definition anyway, difficult targets have no pre-existing binders. So we don't have any frameworks that are known to work. We don't have any light chains that are known to work. We don't have any CDR sequences that are known to work for these particular targets. Usually I'm taking the kind of the hardest case here. So when we say de novo discovery, we mean customer or our own selection chooses a target and we get the sequence or structure, and we have to generate antibody sequences from that and nothing else. Anything else that we would talk about would be optimization. So if you've got something that starts and you want to replace the CDRs or improve the affinity, improve the developability, massively valuable opportunity, we would call that optimization. So what, what matters in the end is getting customers or getting the market what they need, antibodies delivered to specification. Okay, potential benefits of computational antibody discovery. And again, this whole talk now is focused on de novo discovery, so all that to be taken in context. So epitope selection— you'll notice a lot of the things I'm talking about today are echoes of things people have spoken about earlier. I actually think that's a good thing. It means that we've kind of converged on opinions about what the important challenges are. So we and everyone else, um, spoken to lots of people in the industry, and these are the motifs that we keep coming across. So epitope selection is really powerful when you're doing immunisation, and to a certain extent display technologies. You tend to encounter immunodominance. That's the tendency that antibodies that you do find bind to the same epitopes. You get the same epitopes repeatedly, and that can be fine but it can be a problem if you're particularly after an epitope that is not immunodominant. If you can select epitopes and design for epitopes, then you can test out biological hypotheses, and that's extremely powerful. You can come up with ideas about which epitopes are relevant to your disease model, and I think we saw a great example of that earlier with speaker. You can also design in specificity, and you can almost have like a menu, a menu system. If you have got a good computational discovery system, you can say, hey, I want to bind to the Sino homolog, I want to bind to the Murine homolog, and I actually want to avoid this really pesky family member that's really closely related to my target but has got some negative side effects. And you can do that kind of thing. You can also do optimization, which has been a big, heavy topic today, and you can design from the beginning with developability properties in mind, and you can also select the best trade-offs. If you can kind of search the sequence space and look at all the variants and assess them accurately, then you can choose, okay, which things am I willing to give up a little bit on and to improve something else, because usually you can't get the absolute best of everything all in the same molecule. So antibody drug design is a lot about trade-offs. One example is It's often the case that affinity and specificity, you can't get the best of both worlds. And so maybe you want to trade a little bit of affinity in order to improve your specificity or vice versa. And that's a case-dependent basis. And then finally, speed. As I mentioned before, it can be a faster way for cases where it's urgent. So as I said, the middle section of the talk is four sort of angles on the problem or four lenses to look at the problem from a technical perspective. And again, this is with a particular machine learning set of lenses on. Okay, this is the analogy I'm using. I call it the data pyramid. I'm sure there are better ways of describing it. Machine learning is all about data, so data is the first sort of topic. And we can get data on a very specific system. We can take a particular antibody-antigen complex if we have one and we can make mutant scans, we can run variants, we can prod and probe and test that through all kinds of assays and build up a really rich dataset on a specific antibody-antigen complex. That's great, and for specific cases, that's exactly what you want to do, but it doesn't generalize to the more broader set, the type of system you can gather antibody, you can gather data on antibody-antigen interactions in general., and that can be used to infer things about your specific antibody-antigen complex, if you think about it right. You can also gather data on an even broader type of system, so protein-protein interactions in general, and again that can infer your decisions about antibody-antigen interactions. And you can kind of take that as far as you like, you could look at molecular interactions in general, The point I'm making is each layer can inform the layers above, but there's a sort of trade-off. So you can gather data on your specific system, it doesn't generalize well. You can gather data on a general system, it won't be as accurate on a specific system. And in general, at each layer of the pyramid, data at one layer will trump kind of inferences and predictions and guesses from a lower layer. It's fairly obvious, but it's a sort of trade-off. And performance— when I say performance, I mean kind of predictive performance at any given layer— is determined by the data you have at that layer relative to the complexity of the system, and the data in the layers below, kind of weighted by the relevance of those layers to your system in question. So that's one lens, looking at things through the data that's available. And actually, An example, just to make it very specific, you can do transfer learning, it's what it's called in machine learning parlance. If you train a model on protein-protein interactions, that can inform antibody-antigen interactions. If you train a model on antibody-antigen interactions, that can inform your model for your specific complex. Okay, second lens I wanted to talk about was data representations. And this slide is a little bit busy. But essentially, I'm talking about two axes. You can look at your, look at your system in a high level on the y-axis here on the left. You can look at a high level of physical granularity, accuracy, and go for an all-atom representation. Or you can zoom out a little bit and talk about residues, or you can go to a more coarse-grained kind of representation. And then on the x-axis, I'm talking about kind of dimensionality in a simplified way. We can represent a protein, most proteins, as a simple linear sequence, because they're a string of amino acids, typically. We can represent them as surface patches, because after all, surfaces are kind of what are doing the interactions with each other. We can look at 3D structures, because they are after all 3D structures, and they are— it's not just the surface that matters, it's the things underneath the surface that matters. And finally, I've put this on the same axis. Strictly it should be kind of orthogonal to the whole page, but that made it a horrible diagram. We can consider time dynamics confirmations as another aspect. So these are all— all these representations can be dynamic and change over time. And I think the purpose of this slide is just to show you that there are a number of different ways of representing the problem and The representations you choose basically depend on what questions you're asking and what you want to get from the models you're building. In here, I'm not getting into deep learning architectures like graph neural networks, ConvNets, protein language models, et cetera. I'm not getting into hierarchical versus simultaneous generation, et cetera, et cetera. That's kind of getting into the machine learning weeds, if you like. For this talk, I'm keeping it somewhat high level. And apologies to the LIDL lab and the MASSIF authors for using their diagrams without accreditation. It's very bad practice. I couldn't get the references in time, I'm afraid. So, third lens I wanted to look at is, once you've got the data, once you've chosen a representation, what's your computational paradigm? And I see it as a kind of spectrum. We can go from purely modeling without any data sort of taken into account, and that would be a physics simulation classically. At the other extreme, we can talk about looking purely at the data with no kind of model or structure to that. And there I'm just basically talking about statistics. Now, I'm not meaning to denigrate either of those. They're both valuable. In particular, I want to call out statistics as really valuable for a lot of companies. We've noticed it's a really low-hanging fruit. You can get more from your existing experiments by analyzing them in greater depth. So next-generation sequencing and computational power are both pretty cheap these days. If you do a phage display campaign or immunization, you do next-generation sequencing and do analysis of that, you get lots of insights just from simple clustering, looking at trends, looking for motifs. Things like that. So a lot of companies are 100% on that and have been on for some time, but not everyone is there yet. The middle part of this slide though is what I'm most interested in, and in fact when you look at practical modeling systems, kind of pseudo-physics simulation systems, that is molecular dynamics and macromolecular docking, the real-world models, they actually use, have used feedback from experiments to improve, and they have kind of hand-designed force fields and things. They're not based on a pure hypothetical physics model. They are tuned to get better results in the real world. And in fact, if you're doing statistics on your results, then a little bit of domain, domain knowledge gets you a long way in getting in terms of better results. So machine learning again is where we, where we are, and that sits in the middle too. The nice thing about machine learning in my book is that you can combine knowledge with data and. And so there's a lot of expertise built up about antibody discovery and antibody design, and we can combine that with data on, hey, this particular sequence binds to this particular sequence. Great. And that's a really powerful combination. The other thing that machine learning gets us— it's not the only way to get these things, but it's a good one— it gets us descriptions of what's going on. It gets us predictions of what would go on if I did this, or if I had this situation. And it can get us generation of, hey, I want this particular outcome, give me some inputs that might generate that. And that's a really powerful thing to be able to combine all those things in one system. I think that's partly why machine learning is really popular these days. Okay, the final step is you've got your data, you've got your representation, you've got a computational paradigm, In our case, we're building machine learning models. How do you use it? And there's sort of— there's a wide variety of ways you can use it, but when I'm talking about de novo generation, there are kind of two archetypal sort of patterns, as well as a number of other messier things. The first one is, hey, we've got a model, let's just basically replicate high-throughput screening the experimental model of that, let's do it in silico. And it's got strengths. It basically involves taking an antigen and an antibody and making a prediction. Do these things bind or not? Or to what extent do they bind? Or how do they bind? Where do they bind, et cetera? But some kind of prediction. And you repeat that a million, 100 million, few billion times, and out of the antibodies your model predicts are going to bind, you have an enriched rate of binding if your model is any good. The other sort of extreme is what I've called specific generation here. So rather than taking a library approach of antibodies that you kind of churn through at high speed, you can say, nope, we're actually going to generate them one at a time in the context of this specific antigen or this specific epitope. And that's a kind of a different way of doing it, although they are obviously closely related, because in order to generate something that binds, you have to be able to know whether it binds or not. Kind of duals of each other, if you like. Okay, I think those are the four sort of lenses, angles, introduction to machine learning for de novo generation. The challenges in this field are— well, there's many, but a few that I wanted to put out here. Uh, let's just be clear, this is an incredibly challenging field. In 2017, when EndoVerse was founded, we had a great deal of difficulty getting them to give us any money because they said it's impossible. We've tried that several times over the decades. Everyone promises the world and never delivers anything. If you look forward to nowadays, we've got some incredibly powerful systems like AlphaFold, RosettaFold, many other systems. They still don't do a great job on antibodies. Antibodies are particularly challenging, even within protein-protein interactions. And okay, so what are the actual challenges? There's a large search space, as Twika mentioned. We can call it 10 to the 16 as a sort of estimate of human natural antibody repertoire, maybe up to 10 to the 60 if you take sort of more a much wider view of all the potentially possible antibody sequences, and there's a limited amount of data. We're not talking ChatGPT trained on billions and trillions of tokens. We're talking, you know, thousands, tens of thousands of antibody-antigen pairs available. If we're talking about 3D structures, which is the sort of highest resolution data that we get, then we're talking about a few thousand. Basically, so large search space, small amount of data, it's difficult. There is a lack of— and actually, hang on, when I say limited amount of data, it's not just the amount of data that's important, it's the quality as well as the quantity. And when I say quality, I'm talking two things, sort of resolution. We can say this is a binder, this is non-binder. Okay, that's some information. More informative if we get kinetics like on rates and off rates, and if we can get epitope data, that's more powerful again. And then there's also, importantly, there's distribution of data. So for example, there are a lot more antibody-antigen complexes resolved nowadays and antibody sequences published. A large number of them over the last 3 years have been against COVID and the receptor binding domain or the spike protein. Which is great. We've got a lot of detailed information about that particular antigen. So if you want to go out finding and designing antibodies against it, brilliant. But it's very lumpy, if you like. So very related, there's a lack of data on certain target types, and one sort of prime candidate is proteins where the bit of interest is heavily phosphorylated or glycosylated, so post-translational modifications. That is tricky because most protein-protein interaction models are trained on proteins, unsurprisingly, and most representations— thinking back a few slides— don't cope very well with things that are not in the protein chain. So a protein chain is nice and simple. At each position, there's one of 20 amino acids. If you're going adding phosphorylation and glycosylation on at any of those locations, and glycosylation can be extremely diverse, then you're adding a whole lot of complexity. And to make it worse, there is not very much data on those target types. So, um, a further complexification is protein flexibility and dynamics. Antibodies are dynamic, antibodies are flexible, The six binding loops, the CDRs, are generally disordered, which makes things a lot more complicated. The antigen can also be dynamic and flexible. Further tricky part is validation, and throughput is always an issue. If we wanted to make— if I said, okay, can I make this million antibodies, please, exactly as to my description, it would cost an awful lot. It would take an awful long time. And feedback loops are generally slower in this domain than they are in, say, image processing and text processing. If we need to make predictions, so the design-build-test-learn cycle is significantly slower than if you're fully digital because we've got that lab in the loop. It's still powerful, but it's significantly slower. And at the end of the day, difficult targets are difficult for a reason. And the reasons are extremely diverse. There's a lot of biological uncertainty inherent in many of them. It's a challenging domain, and there's a lot of project-specific stuff that can be key. Okay, I don't think I needed to persuade anyone that it was a difficult challenge, but we've gone through a few of the details there. Talking about Antiverse now, our approach is— well, we're a team of 18 a mixture of machine learning engineers, wet lab scientists, advisors. Founded in 2017, headquarters is here in Cardiff, but the team is global. We work with pharmaceutical companies, we have an internal pipeline, and we work with external partners. We've created and validated a number of antibodies, we've done lots of campaigns, we've got our own internal libraries. Our sort of mode of operation is proprietary models and proprietary data, close integration of the lab with computation. One of the things we really focus on is accuracy in order to deliver valuable projects. And as I mentioned, we focus on difficult-to-drug targets. The accuracy to deliver valuable projects is something that's kind of key to everything we do. And one, I kind of wanted to have this slide to talk about strategies. When you've got a low level or no level of accuracy at predicting what's going to happen given a particular antibody antigen, you use naive libraries. Basically, you want to maximise diversity, you want to cover the sort of functional space that you can, you want to maximise developability to save yourselves headaches later on, and you can use AI for all of those things. But it's essentially naive libraries. Excuse me. When you have a good level of predictive accuracy, then you can move to target-specific libraries, and if you use AI-driven libraries designed against the specific target of interest, you can either get away with smaller libraries, or it enables you to make discoveries where you couldn't have done them at all with naive libraries, or you get a larger rate and diversity of hits from those libraries. And that's Antiverse's solution at the moment. We do target-specific libraries. When you have really extreme levels of accuracy, and to be clear, we're not there yet, and I don't believe anyone is, then you can do direct clone design where you say, okay, my model is so brilliant, I'm just going to make this 100 antibodies directly as mabs, and we'll assay them, and then that'll be done. We'll get good results. That is in the future. Okay, given that, antiviruses discovery platform is generally unsurprising. The industry has done a great job of refining this platform, the pipeline. We get a target sequence, we develop a stable cell line with very high receptor expression, we target specific— we do a target-specific library design, we pan, we do next-generation sequencing and clustering, we express, mAbs, and we do functional testing. And as someone else said earlier, I think 6 months is absolutely doable for this kind of thing. Just, you can never make things too simple. We have cells with very high receptor counts to make this possible, and antibody libraries with high specificity, high concentration of antibodies that are binding to that target specifically within each library. Okay. On to results. Where have we got to? We have found we've got binders against GPCRs. Here's one example. This structure is actually not of the GPCR or antibody we have in question. We don't have one for that, but it's an illustration. We found a 40% binder hit rate. I think there was upwards of 80 binders in 20 different sequence families. We got a number of those synthesized, reformatted as mabs, and we did some functional testing. I can't give a great deal of detail about this at the moment because it's subject to a patent application, but we found antagonist function against the GPCR. We had low nanomolar IC50s, which is great news. A significant number of them were very effective. And it's not the only GP cell that we've got binders for either, I should add. I'm aware that I'm running slightly behind time, so I'm just trying to speed up a little bit. I mentioned earlier, if you're trying to do high-throughput sequencing— sorry, high-throughput screening— then you need to have very fast models. We wanted to do high-throughput screening as an experiment, but we found that Structural prediction was high accuracy but not high speed, so we developed our own high-speed antibody structure prediction module. We could predict 16 structures per second, which my best estimate is 500 times faster than the closest competitor, a bodybuilder. The field moves incredibly quickly, so if this is now out of date, please do let me know, but it's my best guess. And that's 16 structures per second on a single machine with an A100 GPU. It means on a single GPU can screen millions of structures per day, and the accuracy on the right here is comparable to competitors. This slide is just an illustration of what I was saying earlier. If we cluster in high-dimensional space, this is the results of a phage deploy campaign we did with a major pharma company. We can pull out sequences with similar properties, and if we find— if we sample from each of these different ones, we know we're covering kind of functional space, We find some of the ones that work well, we can go back into the clusters and pull out some more variants that we expect to have similar functions. Okay, future directions. We want to improve epitope-specific design in terms of specificity, accuracy, resolution. We want to have better understanding of dynamics and conformations, both on the antigen and the antibody side., and AI-driven function prediction would be amazing. For example, we could predict agonism and things like that. It does require— it sort of sounds impossible for things like GP stars. It requires drawing the boundaries of the system much wider. So rather than just looking at the interaction, the epitope and the paratope, and the surrounding residues, you need to kind of draw it all the way through the GPCR through the membrane and include the G proteins on the inside. If you can model that accurately, then you can predict, predict function. That is, let's say, a long-term goal. In terms of scope, we want to tackle the more challenging target types to expand the kind of range of targets we can work in. Again, we're looking to make more therapeutics available to more people, and we want to scale to meet the need. There are lots and lots of targets available We are a small team, and so we're looking to grow. If you have any targets, speaking of which, if you have any targets that are intractable or difficult, please do come and talk to us. We're happy to provide free advice. Get in touch, partnership@zenadverse.io. And thanks, everyone else.

And that's it for me. Matt, thank you very much for that. So in the interest of time, I'm going to ask one question from the audience. So can you elaborate on the evaluation metrics you typically get for a difficult target?

Yeah, sure. So I can distinguish between in silico metrics and practical metrics. The practical metrics are if we want to, if we run a campaign, do we find binders or how many binders, how many functional binders do we get? So that's really the gold standard for us. For difficult targets. So we've run campaigns against two GP cells, and we found binders against both, uh, good binders and a variety of binders against both. So if you like, we could say we've got 100% success rate. I don't expect that to continue forever, um, uh, and I think it was in the order of, yeah, dozens of, uh, significantly diverse binders for both of those. In terms of in silico metrics. It doesn't really matter whether we're talking about difficult binders or not, because the in silico predictions are kind of the same. We typically look at things like— we pose it as a ranking problem. We say, okay, if there's a billion candidate antibodies, can we rank the ones that bind high up in that list? So things like mean average precision, NDCG, etc. That seems to be a lot better than naive metrics like accuracy.

Great, thank you very much, Ben. I think in the interest of time, we're just going to move on to the break. And because we were running slightly behind schedule, we decided to make the break till quarter past, or 10 past, as advertised on the schedule. So thank you very much to all the speakers, Ben, Victor, and Sandeep, in this session. And hopefully see you for the panel in 7 minutes from now, so quarter past.

Thank you very much.