November 10, 2026 | Andrew Bradbury (Specifica) and Ben Hackel (University of Minnesota) pressure-test AI antibody engineering against real measurements on the latest episode of The Chain. They explore where algorithms match results from the bench and where they still fall short. They also trace how better datasets, better blinding, and better developability thinking can make the next generation of antibody discovery more trustworthy and more useful.

Listen in to hear:

• why the field needs prospective, blinded benchmarks instead of marketing claims
• how the Nature Biotechnology antibody challenge is structured across three tasks
• about affinity maturation results where the best AI matches the best experimental outcomes
• why developability screens can overturn “best affinity” winners
• what CDR design gets right today and where it still breaks in practice
• how investor pressure and publication gaps distort perceptions of progress
• the origin of Specifica and the Gen 3 library idea of matched, developable building blocks

Please subscribe to the podcast and give us a rating. It helps other people find and join the conversation. If you've got speaker or topic ideas, we'd love to hear those too. You can send them in a podcast review.

The Chain explores the lives, careers, research, and discoveries of protein engineers and scientists, the impact their work is having on the field, and where the industry is headed. Tune in to stay up to date on the newest advancements and to hear the stories that are impacting the world of biologics.

Links from this episode:
https://www.pegsummit.com/

Host Links
https://cse.umn.edu/cems/ben-hackel

Guest Links
https://www.specifica.bio/

GUEST BIO

Andrew Bradbury, MB BS (MD), PhD, Chief Scientific Officer, Specifica
Andrew Bradbury is Chief Scientific Officer of Specifica. He trained in medicine at the universities of Oxford and London and received his PhD from the university of Cambridge at the MRC Laboratory of Molecular Biology under the guidance of Nobel Laureate, Cesar Milstein. He has worked in the fields of phage and yeast display, library generation, antibody engineering and Next Generation Sequencing for over thirty years. He was a Group Leader at Los Alamos National Laboratory before founding Specifica. Specifica's mission is to enable companies developing therapeutic antibodies with the world’s best antibody discovery platform.

HOST BIO

Benjamin Hackel, Executive Officer in Chemical Engineering, University of Minnesota
Ben Hackel is a Professor of Chemical Engineering and Materials Science at the University of Minnesota. He earned degrees in chemical engineering from the University of Wisconsin (B.S., advised by Eric Shusta) and MIT (PhD, advised by Dane Wittrup) and performed postdoctoral research in the radiology department at Stanford University (Sam Gambhir). Since its inception, the Hackel lab has applied protein engineering technologies to develop physiological molecular targeting agents for molecular diagnostics and targeted therapy, with a focus on oncology and infectious disease.

TRANSCRIPT

Welcome To The Chain

Welcome to The Chain, the podcast exploring the lives, careers, research, and discoveries of protein engineers, scientists, and biotech professionals. We look at the impact their work is having on the field and where the industry is headed. Tune in to stay up to date on the newest advancements and to hear the stories that are impacting the world of biologics.

All right, hello everyone. My name is Ben Hackel. I'm a professor of chemical engineering and material science at the University of Minnesota. It is my great pleasure to be speaking with Andrew Bradbury, founder and CSO of Specifica. Since the days of his medical training at Oxford and London and his PhD work in the Milstein Lab at Cambridge, Dr. Bradbury has made a broad array of significant contributions to antibody engineering. In addition to numerous technological advances, Andrew has been a leader in guiding the direction of the field, including, amongst others, advocating for agent reproducibility, as well as the thoughtful consideration of the role of computational design and analysis and AI in antibody engineering. Andrew, thank you for joining me today. It's been my pleasure.

Why Run An AI Antibody Challenge

So I thought we'd start with the discussion of the recent AI antibody engineering challenge. We can dive into this challenge and all that it has taught the field, as well as using it as a launching point for discussing the state of antibody engineering, including the impact and perhaps the limitations of AI both now and in the future. So the results of this were published in August in Nature Biotech. And briefly speaking, in this challenge, inspired perhaps by the CASP protein structure prediction, organizations were invited to compete in three tasks: affinity maturation, affinity ranking from a selection output, and CDR design. In my take, this was an impressive combination of insight to have this in the first place, rigorous and systematic scientific execution, as well as collaboration amongst many parties with varied perspectives and incentives, perhaps. So I'd love to hear your perspectives on a lot of these elements. One, the motivation and inspiration for the project, two, the interpretation of the results, and third, kind of the vision for the next steps after this. So perhaps let's take those in order. Give me your story of the inspiration and motivation for the project.

So

Building A Fair Prospective Benchmark

it all started really. I was going to a lot of meetings, conferences, all these people standing up saying, you know, AI could design antibodies, they could do this, they could do that. It seemed to me like, well, experimental scientists were going to go out of business. There was no need for them. But that said, I am old enough now to have participated in a few hype cycles in antibody engineering. And I think I personally contributed quite a lot to phage display as being, you know, the answer to all things antibody discovery-wise and how a phage display was going to stop us immunizing animals. That proved not to be the case. And so I what I thought we needed was really a prospective study to understand really what AI could do. Because the other thing I'd noticed was that a lot of people were talking about AI and how great it was for antibody engineering, but nobody was published in peer-reviewed papers. And that remains quite true now as well. So what what where do you read about it? Well, you read it on white papers on company websites. And so this just made me a bit queasy, let's say. So we had during COVID, we had generated really large data sets on antibodies against RBD from our semi-synthetic libraries. And so speaking to people internally, Frank Erasmus and others, we thought, wouldn't it be good if we provided some of these data sets to the community and asked people that were doing AI to undergo some challenges, the ones that you just described. And so I reached out to the editor of Nature Biotech and asked if she would be interested in an announcement manuscript of this kind. And to my surprise, she said, yes, but only if you promise that you'll publish the results in Nature Biotech. So that was great. And so we we got to work. Turned out it was a lot of work, I have to say. So the first thing we did was we we accumulated a whole bunch of CROs, asked them that they would donate effort to undertake the analysis. So have to really thank the the participating CROs. Azenta was responsible for making the antibodies. And they they didn't do it for free, but they did at a significantly discounted rate. And then once the antibodies were were made, they were sent to Carterra, and some of them were also sent to Sapidyne for affinity measurement, and they were also sent to Mosaic as well for developability analysis. And specifically, of course, provided the data sets. So what I because I just really did not know what was true and what was hype, I decided we would make it as easy as possible. If you look for the number of publications, then RBD is the most studied protein in history. I think Rubisco is behind, but RBD is ahead. And then we wanted to use also things that were going to be helpful within the context of a true antibody discovery pipeline rather than more esoteric things which were perhaps interestingly scientifically, but not necessarily useful in a commercial context. So we had published a method of doing affinity maturation. This involved creating libraries using fixed CDRs, perfect sequence viabilities in HCDR1 and HCDR2, and then separate libraries in L CDR1, L CDR2, and finally L C D R3. We tended to keep HCDR3 constant when we were finished maturing, and the reason was to not get epitope drift. And so we provided three data sets from those where we had selected binders and then asked participants to create using that, using those sequences, create an antibodies that would have the highest affinity. And we'd also done the experiment, we'd done the real experiment using a youth display. We haven't, we we did not share the data of that. That was that was that was to be done later, because we wanted this all to be done blindly. And it was interesting. The the best the best AI algorithms were as good as the best experimental. I think the improvement was, I can't remember exactly maybe a couple of hundredfold, going down to I think just under 100 picomolar in the case of the AI, and just slightly above the 100 picomolar and the best experimental, but they were actually statistically indistinguishable. Now, so there was one company that was Aureka that did extremely well. But there were an awful lot of other companies that, and not just companies, also I don't want to just blame companies, also academic groups that submitted sequences, and many of them just didn't bind at all. And there was a whole bunch of them that in fact were made had worse affinities than the starting parental. And there were others that sort of on this on this on its spectrum going from famous parentals pretty close to to well Aureka was the best, and then the others were sort of behind that. Interestingly, there was one group that didn't use AI, that just did a statistical analysis and made a consensus sequence, and they were actually, I think, second or third all altogether. So you didn't have to use statistical, you didn't have to use AI to get to get great results. Now, this was just a single antibody, and so we don't know how transferable these results were. And in fact, internally in Specifica, we are working on trying to see how how broadly applicable these methods are. We in fact did participate in in the competition. We didn't put our results in because we just didn't think it was fair, it would have seemed like a conflict. And and so now what we're interested in is really I think using AI within the context of a true pipeline to see whether we can we can we can improve speed and or and or functionality and/or affinity. So that was the first challenge.

Affinity Maturation Results And Surprises

Now the next challenge, I would argue, is actually rather more broad, broadly applicable or broadly of interest, in the sense that the first channel was our way of doing affinity maturation. Not everybody does it like this, but the second challenge, second, second challenge was something that everybody comes across. You do you do an immunization, you do an in vitro discovery, you do NGS, and you end up with this massive list of sequences. And so, how do you pick sequences from that? And secondly, everybody will have noticed that many, many sequences in a selection output are actually clusters of HCDR3s that are quite similar to one another. And so within the context of a cluster, how can you identify the highest affinity antibodies? And so we had previously shown that if you took a selection output and you clustered it based on the HCDR3 sequences, then you took the most abundant antibody that was always a binder. But it was not necessarily the best binder. And in fact, something like 30 to 35% of sequences that were in a particular cluster had higher affinities. So you just pick random, had higher affinities than the most abundant cluster. And so what I wanted to see, or we wanted to see, Frank and I, was can AI improve upon that? And so we had identified, this was again the RBD data set. We did a selection output on that. We had some 30,000 sequences. We identified three clusters, the most common clusters or the most abundant clusters, which had 400, 2,500, about 3,500 different sequences within them, all clustered based on the HCDR3 sequence. And we provided the affinities of about 140, I think, of those. And as it turned out, this wasn't done deliberately, but it was, it turned out to be very fortuitous. One of those clusters had antibodies that were extremely developable. One of them had antibodies that were not very developable, and the third was somewhere in between. And so what participants were asked to do was identify high-affinity antibodies in these three different clusters. And so they had, as I said, 400, 2000, and 3.5,000 sequences to choose from. And also, you have to identify that the antibodies that were produced had to be developable. And it turns out that there wasn't really any the the three clusters, the developability of those clusters were maintained in the AI-dependent PICs, if you like. So, in other words, AI didn't appear to be particularly good at identifying in this particular data set the developable clusters. But then the real shock, when we sort of analyzed the data with one exception, which is the University of Washington, they were able to pick the highest affinity antibodies from two clusters, although in the third cluster they came out sixteenths, so they they did it well for two, but the third they didn't do terribly well. That was a real shock. I mean, I thought that this was gonna be a really easy, easy challenge. You know, you have you have two and a half thousand sequences, they're all clustered on the base of HCDR3, they're all similar, you've got a bunch of affinities in there. You know, picking the highest affinity I thought was gonna be really easy, and it turned out not to be the case. And then the third challenge was taking this set of 30,000 different sequences, all binding RBD, with the caveat, there may be sequence errors or PCR errors in there, so there may be some non-binders. And also the library was a single-chain library, but we converted to IgG. So again, there may be some conversion issues there as well. So with those caveats in mind, the idea was design CDRs that were not in the library. So, in other words, design antibodies with novel CDRs not in the library and see if they bound. And that was very interesting. So Zencor, which doesn't consider itself an AI company, they think they're basically an antibody company, they're good at that. They were able to design an antibody with an affinity of three picomol, where the best coming out of the out of the library was nine picamol. So the nine picomole was taking all those all those sequences straight out of selection and just measuring the affinities of some of those, including some of the ones that have been identified. The best affinity we got straight out of the library is nine picomol. Zenko was able to design one antibody with a just under three picomolar affinity. However, it didn't elude from the HIC column and it had some weird kinetic behavior. So it probably wouldn't have been, you probably wouldn't have taken it forward if you're interested in developing an antibody sort of therapeutically. But nevertheless, the affinity was remarkable,

Why Clone Picking Stayed Hard

excluding all the others that failed the developability. Then the next antibody that had been designed that was developable had an affinity which was equivalent to the best, the best experimental affinities, about nine pico molar. So my conclusion from all of this was that it's likely that AI may be helpful in affinity maturation. At the moment, it doesn't seem to be terribly good for clone picking. And AI design, I think we need we need to see some more data. Now, people will say, well, that was in a year and a half ago, which is true. So that because this we we we we announced the competition in, I think it was two years ago now, and then in January of last year is when we provided the data sets. People people provided their sequences within a couple of weeks. The idea being to avoid any any possibility of doing experimental. We wanted people to really use their AI algorithms. And then it took time for Azenta to make the antibodies and have have them all go through the through the system. And then we had to repeat some and and and I mean, these things take time and an incredible amount of work. So there is that. I think now if we did it again, who knows? Maybe the results would be better. And we're we're quite interested in finding out about that. The other thing that people will say is that some of the big names in AI antibody discovery didn't participate. That's true. Part of the issue was, and I I think it's valid of valid objections. It wasn't the submissions were not informatically blinded. And so we we could see where the different companies were performing. And I I understand if your whole business is based on using AI to do something and and somebody's able to say, well, you know, you're not doing it terribly well. It was basically our integrity that allowed us to keep to sort of keep the blinding clear. In the next competition, that blinding will be done informatically, so nobody knows, and it'll be sort of like a double blind. Nobody will know which which antibodies came from where, except when the unblinding occurs of the winners. So one of the conditions from Nature Biotech was that the winners had to be unblinded and they had to provide details on the on their on their code. And most people were fine with that. There wasn't any issues. So that's the competition. Now what's next? The next competition is going to be, it's not going to be organized by Frank and I. The Antibody Society is looking after it. And so if anybody's interested, please reach out to me or to the Antibody Society. We are we are going to be helping with that. But it's interesting, since this has come out, a lot of CROs are interested in participating. So there's there's been a lot of interest in providing additional additional measurements. And I think this is likely to become a forum actually for also comparing developability measurements, affinity measurements, all these. So in addition to the to the AI and the antibody discovery portion of things, there's also a sort of background going on of comparison to different methods. But unlike this first one, the idea is to throw it open, give people four months now to get antibodies, and you can get them anywhere you want. So use in vitro libraries, immunize transgenic mice, use AI, design test build cycles. So the idea is to see how how these how these different methods compare within a real timeline. So to turn it into something that's really valuable within the context of a pipeline. So the Antibody Society is identifying targets. I know nothing about it, and I've I don't want to know. Uh, we would like to participate as one of the one of the one of the competitors now this time. And we'll we'll see, we'll see, we'll see what happens.

That sounds very exciting, I yeah. Excellent, excellent summary of of the prior challenge and very much looking forward to the to the next one as well. Numerous questions to follow up.

Are you enjoying the conversation? We'd love to hear from you. Please subscribe to the podcast and give us a rating. It helps other people find and join the conversation. If you've got speaker or topic ideas, we'd love to hear those too. You can send them in a podcast review.

You very nicely outlined that in many cases, you know, experiment was comparable or even superior to the outcomes at the same time. You, I think, quite fairly acknowledged some of the perhaps limitations, be it this was some time ago, particular target, etc. I'm curious, maybe from an optimistic perspective or an opportunities perspective, where do you think the highest optimism should be for the outcomes looking perhaps even stronger than next time? Is it simply time has passed and the algorithms have gotten better? Is it the data sets are getting larger? Or other things. I don't want to keep one on the list and put words in your mouth. So what what yeah, where do you think the next biggest advances are? I guess is really ultimately what I'm driving for.

Right. So I think so I spoke to Charlotte Dean about this. She's a professor at Oxford and

Novel CDR Design And Developability

she's done a lot on AI. She doesn't think that there has been a significant increase in data to allow for a significant improvement of algorithms. And yet there do appear to be a lot of people standing up and talking about zero-shot design, where they're getting 10% of their designs are actually binding, and some with what appear to be quite quite good affinities. So I again I just think that we have to we have to do another prospective type experiment in order to understand whether this is true or not. You know, I hate to say this, and and and it was sort of subdued a bit in the publication, but we mustn't forget that many of these companies are funded by investors and they're funded with vast amounts of money. And so there is a there is a there is a desire to demonstrate that they're working. At the same time, I would also argue that if I was an investor in in these companies, or even a scientist, I'd like to know how well the company I'm investing in is actually performing in the real world. And I think there has been a push in the AI community to identify what is the most, if you like, sexy challenge, which is designing antibodies de novo. Antibody discovery, I think, is pretty much a solved problem. We have we have we have in vitro display, we have immunization, and generally we can get antibodies against these different targets. I think AI will become an another way of perhaps designing antibodies, but ultimately I think it's going to be much, much more valuable where you have to sort of improve improve pipelines and improve antibodies rather than de novo design. I think de novo design is just a very, very sexy sort of technology, and that's what everybody's getting very excited about.

Well, that makes me think about the non-antibody world as well. Certainly, many, many topologies and formats, both antibody and non-antibody alike, I guess. So antibody but unique formats and completely non-antibody topologies. I guess what are your thoughts in that space? Certainly a lot of the de novo techniques, and that's kind of why it came to mind when you made that comment, yield very non-immunoglobulin type folds, some alpha, some beta, certainly a preference towards alpha predominantly, but emerging betas as well. Do you think those are better suited for success than antibodies or not? Do you think a non-antibody challenge would perform better than antibodies given the entropic cost of binding concerns? So plenty of things to think about there.

Right. So, I mean, I had I when I was younger, I was quite interested in looking at other doing a bit like what you've done. We we we fragmented some genomes and did some selections look using ampicillin to filter out open reading frames and beta-lactamase, sorry. And and I was surprised, I remember when we did those experiments that beta- lactamase was able to filter out coherent folding units, if you like. It didn't make sense to me until I spoke to Jeff Waldo, who had developed the GFP folding reporter system, and and

Next Competition And Better Blinding

he said, Oh no, it's it's obvious. Beta-lactamase was behaving as a folding reporter here. And I think that makes perfect sense. And so I think the idea of making fragments, I think is great. I just think that antibodies are so entrenched in the therapeutic landscape. Everybody loves them, everybody understands them, you know how to make them. To produce something else that just does the same as an antibody, which binds, is binding. You know, why would you want to do that unless it had some intrinsic advantages? And apart from things like maybe cost of goods and small, small size, I don't see the in advantages, potential advantages over traditional antibodies. And I think this is more a I'm sure that some of these small fragments could be extremely good in the therapeutic context. Some of them might be hemogenic, others not. But I think when something is working as well as antibodies and bi-specifics and all the other variants thereof, I think it's difficult to break in with the new technology.

Yeah, I think it's a fair perspective all respectfully assert that I think the small size and modularity has certainly some advantages in several applications, but that can be a discussion. For another day. What I want to get to next, I think, is kind of continuing that vein, but taking perhaps a step back, broadest perspective beyond the antibody challenge that you that you led, would be looking to the future, five years, ten years out perhaps. What do you view as the most influential and addressable challenge currently facing the field, be it experimentally based or or and or AI? Now you said another way, what should be the most impactful advance that we'll see in the next decade?

I can tell you the problems that I think we need to solve. Yeah. Those are binding to difficult targets. Um GPCRs, IM channels, transporters, things like that. There are a lot of experimental methods that work, but they all seem to be sort of one-offs. And so it's not clear whether they're whether they're broadly applicable or if they're just they just happen to work by chance in that one particular experiment. So a routine way of doing that, I think, is something that really would be, would be, would be of interest. And is something that there is there are a couple of AI companies, I can't remember which one it was now that where they designed first a GPCR that was soluble by playing around with the with the transmembrane domains and then generated antibodies against that. And those antibodies were able to bind cells. This was a meeting. I don't think that's been published. I haven't seen that. But anyway, that that's certainly promising. And then I think the other thing that is really going to be challenging, and again, I would like to be proved wrong, but I suspect it's going to be brute force. And that is which combinations of binders in which orientation and structure are going to be the best by specifics in order to carry out a particular biological effect. Right now, it really does seem to be very much a massive screening effort. And if if that could be if that could be really improved using using AI of some kind, that would be great. But right now, you see there are many different ways of making biospecifics. People get different binders, they 10 different binders, put in 10 times 10 different combinations in two or three different orientations. I mean, before you know it, you're talking about hundreds of hundreds or even thousands of different different sorts of sorts of constructs, each of which needs to be tested, produced, and then tested. I think those are the two big things.

In that latter case, no, my apologies. In that latter case, is your intuition

The Next Decade Hard Targets

such that in all those combinations, the winners win because they get the geometry right, because they get the epitope right, because they're driving the right receptor geometry after binding? Is it something else? Or is it different answers in every different use case? And I asked that, because that I think is what's going to drive which technology or insight is actually gonna rule the day and solve the problem, is us figuring out what the real driver of the solution is, if that makes sense.

Yeah, I fear it's gonna be the latter. I think each each each each problem will have its own particular solution, which is which is why I think there are these massive screening efforts. And I think it's gonna be challenging.

Fair enough. So I want to make sure we have some time to get to one other topic, and that is on Specifica. So you certainly highlighted very nicely the level of collaboration that went into the AI antibody challenge, and in no small part, Specifica, and you and your colleagues led the way along the way. I'd like to hear a broader story, perhaps, on you know, the origin of Specifica and the journey along the way, and and from a lens of how can the broader community, from an antibody engineering perspective, learn some lessons.

Right. Well, so I founded Specifica with Ken Sharples, who's my business partner. And he, I'd always wanted to set up a biotech company. And I'd done a few in the past which weren't successful, where I'd been I I just remained an academic and I hadn't jumped ship. And then Ken reached out to me after he'd been involved in something else and said, you know, would you like to would you like to set up a biotech, biotech company? And I at that time the kids had left home, so empty nest, we didn't have to worry about about university fees and the like. And it was I was able to, I could retire from Los Alamos. I had a had a bit of a bit of a little bit of a pension, which was enough to keep me going if things went badly. And so and so we agreed and we we went ahead. And at the beginning, we we'd already had always had this idea that we could do antibody discovery. We made a few libraries, and then really I would say we bet the whole company on this this Gen 3 concept, as we called it. And the only reason we called it Gen 3 was because we didn't didn't really have a have a good idea of what else to call it. And the the concept behind it was to make an antibody library where every antibody in the library could be a real antibody. And so we had a few principles, start off with developable therapeutic antibodies. So we figured that pharmaceutical companies on the whole will have done all the work to ensure that their antibodies were developable. And so this was just after Jane had come, Jane Natal had come out from Adimab, and where they looked at 130 odd different therapeutic antibodies. So we took, I took that data and I sort of rearranged it and I identified a bunch of different therapeutic antibodies that that were developable and were also diverse at the germline gene level. And then we did some NGS of lymphocytes, so B cells, and were able to identify CDRs. And the idea was that you only put CDRs like

Specifica’s Gen 3 Library Story

VH3 family CDRs into the VH3 scaffold. So then that way everything should be working from a folding point of view. And then we poached all the sequence liabilities. I mean, it's quite a straightforward concept, really. And then for the HCDR3, because we couldn't build that many HCDR3s, synthetic capability still isn't really up there. Uh, we just took 10 donors and amplified the HCDR3s and put them in. And to our surprise, it worked surprisingly well. What we were trying to solve was really the developability issue, because we thought we start with developable cast scaffolds, we put in developable CDRs. But it turned out that the affinities that were coming out of this libraries were remarkable. And I remember when people came to me and said, Oh, well, it looks like we've got sub-nanomolar affinities here. And I said, you know, go back and measure it again. You know, I just didn't believe it because all my previous attempts at making antibodies from libraries at had really if we got a if we got a single-digit nanomolar, that was fantastic. And here we were routinely getting sub-nanomolar. And we realized that this actually was a platform that stumbled in a way across a formula, if you like, that just worked very well. I think in parallel to the scientific innovation was also a business innovation in the sense that we were now offering those libraries without any royalties or milestones, which wasn't the case for our competitors. And we also did discovery campaigns without royalties and milestones. And when we started, I mean, we really we really went the extra mile. We, you know, if somebody came to us, asked us to do something, and we failed first time, we just keep going, even if it cost us money until we eventually succeeded. And I mean, fortunately, we we we don't fail that often anymore, so we normally can deliver. And when we can't deliver, it's usually where we say up front, look, this is a really difficult experiment. I'm not sure this is going to work. But that that that that that sort of slowly built up. We were able, we were selling libraries, we became more and more well known. Now we've sold libraries to over 50% of the top 20 pharma by market cap. We we we do something like 60 discovery campaigns a year. So we're we're we're we're really doing doing doing well, I would say. And now we were acquired four years ago by IQVIA, and we now have an RD budget where they we are trying to develop new selection strategies and also new library formats. And that's that that's that's what I enjoy. Um so that's that's that's the the way it's going. I must say personally, I much prefer the I didn't realize I was didn't think I would say this, but I much prefer being in the commercial world than in the academic world. I got so sick of writing NIH grants that were that never got funded, it was just so painful. It felt so so futile. And I don't think it's got easier since I left.

I want to come back to something in the middle. I'll leave that last bit for now. Come back to something in the middle. You commented on this what you said, surprising sub nanomolar binders after you took the Jane and et al. paper and focused on some developability. So, you know, under the umbrella, I guess, of developability begets innovability or evolvability, is is that kind of consistent with what you saw? And if so, do you attribute that then to just a larger fraction of the library was folded? So the functional library size was so large that with a bigger distribution, the tails just had a chance to get out wider. And or is it more coming back to entropy again that the molecules were better folded and so you wouldn't paint that entropic crossed upon binding? Some of both, neither.

So I yeah, so I think you've done a good analysis of what was going on. So I think by ensuring that the CDRs were matched to the frameworks, we were ensuring that only well-folded CDRs were embedded. So I think the the molecules folded much better. We were also very, very careful building the libraries. I mean, we've now built over over 250, but I mean building libraries is a challenge, you know, you know. And and I think it's it's unfortunately it's it's concentrating on on the on the on the small details that I think make the ultimate difference in in the quality of the library. And when we when we when we deliver a library, we have a massive quality control sheet where we provide all sorts of stuff like diversity and and tonal dominance. I mean, it's it's very, very deeply characterized. I think I think by concentrating on developability, we were inadvertently concentrated concentrating on CDRs within scaffolds that folded well. And I think that's why. So yeah, I guess you could say we it was it was perhaps the yeah, the entropic aspect of it. Um but I think it's also just good folding. And and the immune system has evolved over a long time. And the CDRs that we put in, we ensured every CDR has to be found in at least three data sets. So we know that these are not uncommon weird CDRs. They're all that there's CDRs that appear frequently. So yeah, I think it's just a matter of good folding and and and and good quality antibodies and and a greater functional functional percentage. I think that's very important. Yeah.

So a combination of all. Excellent. Well, I have another dozen things I'd be happy to talk with you about. I want to stay within twofold of the time I promised you. So I'll give you one last opportunity. If there's anything we didn't chat about that you want to, go ahead. If not, I'll

Final Thoughts And Call To Participate

thank you for your time and really enjoyed the conversation.

Well, I think the only thing I'd like to say is if you're an AI company or a group out there doing AI, please participate in the next competition. Your success or failure will not be revealed to the rest of the world unless you do do extremely well. And I think it would be, I think that's the only way the field is going to advance. If any everybody who knows about this stuff and is doing this stuff participates and works together, and together as a field, we can really move ahead. So with that, thanks very much, Ben.

Oh, my pleasure. Really enjoyed the time. Thanks so much. Thanks.