Scientist Stories: Gregory Winter, Pioneer in Directed Evolution of Antibodies

Sir Gregory P. Winter was born 1951 in Leicester, United Kingdom. He got his PhD 1976 at the University of Cambridge, the first Cambridge, and he's now a research leader emeritus at the MRC Laboratory of Molecular Biology in Cambridge. He realized early on that phage display could become a very useful method when you work with antibodies. This would be a tool to derive very high-affinity antibodies for any target that you wish. With strong determination, he spent all the effort required to make it possible to display antibodies or antibody fragments in functional form on the surface of the filamentous phage so that he could then fish out the good members from libraries. Sir Gregory Winter, I now welcome you on stage.

Thank you. Well, good morning, ladies and gentlemen. It's still just the morning. Sarah, thank you for your introduction. May I also thank you and your colleagues on the Chemistry Committee and their advisors for all the hard work they had to do that ended up with me getting a share in this prize. So thanks. I would first like to acknowledge and thank my family. It gives me great pleasure they're all here today and will have to listen to me for a whole half an hour without interrupting me or saying, "Dad, we didn't want a lecture on it." So— I would also like to thank the academic institutions— that I— oh dear, now this is going to start messing about— to thank the various academic institutions who've supported me throughout my scientific life, in particular the MRC, the Medical Research Council's Laboratory of Molecular Biology in Cambridge, the Medical Research Council's Centre for Protein Engineering in Cambridge, and Trinity College, Cambridge. All the same Cambridges. I would also like to acknowledge the key role of a startup company, Cambridge Antibody Technology—fourth time I've mentioned Cambridge—in the work that I'm going to describe. I'd also like to thank some individuals—I hope they recognize themselves—from some years back. I'd particularly like to thank John McCafferty, who who was the first to show the display of functional antibody fragments on phage. I'd also like to thank Alan Furcht for his hand in a number of things, but particularly when he was director of the Center for Protein Engineering, for accommodating the greater part of my group as it was developing the phage technology. And all the others on the slide, who are probably the most important because they did the work. They did the work at different times. They all had key roles, and they are the anonymous "we" that I shall refer to during the talk. Finally, I would like to acknowledge Richard Lerner and his group at the Scripps Research Institute. Although we were certainly scientific competitors at that time, their activities complemented our own, and actually confirmed our own faith in our work. It actually helped us persevere through many technical difficulties, because if we— we knew that if we didn't manage to do it, they would. So first of all, you've had a tutorial on phages. I'm now going to give you a tutorial on antibodies. Antibodies are part of our immune defense system. They defend us against infectious agents such as viruses and bacteria. They're raised by the immune system in response to infection or vaccination. An IgG antibody, as shown on this slide, is a large Y-shaped molecule. It's got two arms and a stem. It comprises two heavy and two light chains of linked protein domains. In this case, it's the white and the red. The heavy and light chain variable domains, which are at the end of the arms, come together to form a protein scaffold of beta sheet, which I've marked here. That's surmounted by six loops— these red loops numbered 1 to 6— which are responsible for binding to the antigen. And in binding to the antigen, they can block the process of infection. In addition, the antibody stem here is able to recruit the immune system to kill the infectious agent to which it's attached. Now, the variability of these antigen-binding loops is created through rearrangement of different germline segments of DNA during B-cell development. In heavy chains, it's created through random combinations of 3 segments, the so-called V, H, D, and J segments, as marked on the slide, each of which has multiple members. So you get tremendous combinatorial diversity by throwing these germline segments together. In the light chain, it's created by random combination of 2 sets of gene segments. In addition, where they join, there's editing takes place, and further diversity can occur during during the DNA rearrangement process, as well as random pairing between the heavy and light chain domains as the heavy and light chains are rearranged independently. So after DNA rearrangement, each B cell, which I've got here— so this is before rearrangement, after rearrangement— it encodes a unique antibody on its surface. So the rearrangement process leads to expression of antibody on the surface of the B cell. As the DNA rearrangement's proceeding in parallel and independently in many B cells, this creates a library of B cells, each with a different antibody on its surface. If the displayed antibody encounters an antigen, the B cell is stimulated to differentiate. And it can do two things. It can differentiate to become a plasma cell, which essentially is a factory for production of soluble antibody against the cognate antigen that that antibody has recognized. Alternatively, it can become a memory cell, as marked here, where the antibody genes are subjected to random mutation within the cell. On further encounter with the antigen, the memory B cells compete for antigen, and those displaying mutant antibodies, which have got the highest binding affinity, are favored for further rounds of differentiation. This leads to an antibody response that improves with restimulation with antigen. It's a process known as affinity maturation. So in summary, the immune system is a simple evolutionary system. At its heart is the B cell. It's a phenotype, a genotype firmly linked. It's essentially a genetic display package, with the antibody displayed on the outside of the cell to encounter antigen, and the genes encoding the antibody within. Now, although nature developed antibodies to protect against infectious disease, man has further developed and evolved them for the treatment of noninfectious disease, such as inflammatory disorders and cancer. Man-made antibodies can be used, for example, to block the biology of protein receptors and ligands, such as are involved in inflammatory disorders or cell growth or even T-cell activation, or they can be used to kill target cells by recruiting the immune system to attack. It's the development of antibodies for treatment of noninfectious disease that's revolutionized the pharmaceutical industry. And what this slide shows is the top-selling drugs in 2016. Previously, the pharmaceutical industry was dominated by chemicals, but what we can see here— I've marked in red among those top 10 those antibody products. So actually, the pharmaceutical industry in the top-selling products is now dominated by antibodies. We've got 6 antibodies, and the number 1 antibody is an antibody called Humira, which I will come back to later. These antibodies are mainly used for treatment of autoimmune inflammatory diseases and cancer. The technologies to make such antibody pharmaceuticals required several years of inventions and developments. It started with the invention of hybridomas by Cezar Milstein, actually also at the MRC Laboratory of Molecular Biology in Cambridge, in 1995. They immunized mice with target antigens, such as human cancer cells, The spleens were harvested, and the responding B cells were immortalized by cell fusion. The hybrid cells, or hybridomas, were then screened to identify those making monoclonal antibodies against the target antigens. Although this technology generated many useful research reagents against human proteins and cells, the mouse monoclonal antibodies were seen as foreign, and when injected into patients, this compromised the use. In fact, it could lead to life-threatening reactions. At that time, attempts to make human monoclonal antibodies against cancer cells proved impossible, not least because the immune system has tolerance mechanisms that prevent it making antibodies against self. Generally, this is fortunate. Where these mechanisms break down, it generates terrible autoimmune disease. By the mid-1990s, solutions for making antibody therapeutics began to emerge through the application of protein engineering. This is a branch of genetic engineering in which the genes of a protein are altered and the encoded proteins are expressed in a host cell. The first solutions here involved creating what we call simple chimeric antibodies, in which we take the entire antigen-binding regions at the end of the mouse monoclonal antibody that, for example, could be against a cancer cell, and we transplant the entire region into a human antibody. So the human antibody is marked white. So this antibody is two-thirds human, one-third mouse, and should therefore perhaps be better tolerated when put into patients. Then came humanized antibodies, down here, in which the antigen-binding loops were grafted from the mouse scaffold— this beta-sheet scaffold with the 6 loops— were grafted from the scaffold to its human counterpart. Now, as the inventors of humanized antibodies, which are up to 95% human in origin, We argued that such antibodies might in practice be regarded as synthetic human antibodies, as in any case, these loop regions differ between one human antibody to the next. Experience has actually borne this out. Humanized antibodies are well tolerated in the clinic and account for the majority of antibody pharmaceuticals on the market. But at the time, we weren't actually so sure that the human immune system wouldn't somehow sniff out the taint of mouse that's in those loops, and it might treat them as foreign. You know, I have a very anthropomorphic view of molecules. I think you need to think about them like that. And I was very uneasy about it, despite my logical arguments that they were just the same as human antibodies. So we were very receptive to an opportunity to make fully human antibodies that emerged from a technical advance. One of the rate-limiting steps in making chimeric antibodies— these are the 1/3 human, 2/3 mouse— was the isolation from the hybridomas of the genes encoding the antibody variable regions. In theory, this is straightforward. In practice, actually, it's a pain in the neck. And to simplify matters, we thought we might be able to use the polymerase chain reaction to amplify and clone the genes from the hybridomas. Now, in the polymerase chain reaction, target regions of DNA— so I've marked one here— can be amplified by repeated cycles of extension of two flanking primers by DNA polymerase. And by repeating the cycles, one can get a dramatic amplification of that DNA and clone it. But since we didn't know what the sequences of the genes were in any particular hybridoma, it wasn't clear that we could design PCR primers that would work on any hybridoma. What we did was to compare lots of different antibody sequences, and we spotted regions at the ends of— so if we look here, this is a heavy chain sequences. We'll look at nucleotide conservation at the end. We can see regions which are 100% conserved with regions where they're not conserved. And we thought that maybe by playing around with experimental conditions, we might be able to find a set of PCR primers that could amplify any mouse V gene from a hybridoma. So we set ourselves that task. It involved a lot of fiddling about. It was the early days of PCR. But in fact, we did find a set of primers that worked for most hybridomas. We could actually amplify the genes from most hybridomas we were given. And we later developed a similar set of PCR primers for amplifying human antibody genes. These primers made it so much easier to clone antibody V genes, and it is that that presented the new opportunities. For example, instead of isolating the mouse V genes from antibodies from hybridomas, why not amplify them directly from the spleen of an immunized mouse? If we expressed the encoded antibodies, perhaps we could find the ones with antigen-binding activities. This offered a simple alternative to hybridoma technology for making mouse monoclonal antibodies. But better still, if we could make libraries of human antibodies, we might be able to bypass immunization and find ones with binding activities to human self-antigens as we need for therapeutic use. So we were excited by the potential, but first we had to find an expression host. We looked at bacteria. It had recently been shown that functional antibody fragments could be expressed in the bacterial periplasm. We took libraries of antibody fragments from— well, we took libraries of antibody genes and expressed these as fragments in bacteria. Those genes came from immunized mice. And we soon discovered fragments with antigen-binding activity. Now, by this time, we'd become aware of a similar approach being explored by the group of Richard Lerner at the Scripps Research Institute, who had similar results. But both of our groups were completely unable to detect antigen binding without the use of immunization. So clearly, our screening methods weren't really powerful enough, certainly if we wanted to make human antibodies. We therefore looked back at how the immune system manages to select antibodies. As I explained earlier, at its heart is the B cell. Could we find a B cell mimic? We thought of many possibilities: display on mammalian cells, display on bacteria, or even display on bacterial viruses. But it looked like I would need some more staff, but I didn't have the money or the space to accommodate them. At that time, I met David Chiswell and John McCafferty from Amersham International, who, due to an error in the counting processes, were in the process of being made redundant. They were excited about the possibility— actually, any job, really, they were excited about the possibility— but they were particularly excited by the potential of the library technology, and we wondered about forming a startup company to do that. But it was a struggle. To get funding, but finally I succeeded through an old friend, Geoffrey Grigg. He'd set up a startup company, PepTech, in Australia that was mainly backed by the moms and dads of Australia, but it also seemed to be backed by a— I would call the horse racing fraternity. I had to go and meet the shareholders on a yacht in Sydney Harbor, and I remember one of them saying, "Oh, I'll give Greg the money." let's see how the boffin trots, which was— so Peptek, in fact, in the end became the major shareholder in Cambridge Antibody Technology. David Chiswell and John McCafferty joined. And John came to work in the laboratory of molecular biology. And we decided to focus our energies on the beast you've heard about in the previous talk, the filamentous bacteriophage. So some 4 years earlier, George Smith had shown that filamentous bacteriophage was capable of displaying peptides, and that the phages bearing the peptide epitopes on the phage gene 3 protein could be selected by binding to cognate antibodies. So could this phage serve as our mimic of the B cell? Instead of peptides, could we display folded antibody fragments on the phage? And in this slide, I've actually used a different representation of phage. This is what I call our Dalek representation, which I much prefer to the very long phages. This is like something out of Doctor Who, and we rather liked it, so we've been using this very simplified representation. So what we have here is the antibody genes on the inside of the phage and the antibodies, the antibody fragments, so-called single-chain FV fragments, where we have a linked heavy and light chain variable domain on the outside. They're fused to the phage through the gene 3 protein. So then we tested the recombinant phage for binding to the, the Henegg-White lysosome, which was the antibody corresponding to these V genes. We'd taken a monoclonal antibody, a hybridoma, and we'd cloned it in here. And we found out these phages, using ELISA assay, bound very selectively and specifically to hen egg white lysozyme, but not, for example, to turkey lysozymes. They're very specific, and everything looked very good. So we now mix the phage antibody with a large excess of wild-type phage and attempted to enrich the phage antibody to show the— demonstrate the other work of George Smith and Parmley, the affinity selection process. And we found that we could select this phage by multiple rounds of affinity columns. In fact, on one pass, we could enrich the phage, our antibody phage, 1,000-fold. And by growing the enriched phage and passing it down the column a second time, we can enrich it a further 1,000-fold, which means we've got a million-fold enrichment over 2 rounds of selection. And in principle, using 3 rounds of selection, you could get a billion-fold. And then you could go up to George's gazillion if you just increase the number of of cycles. So we were very pleased, really, that these folded fragments could express on phage, because it certainly wasn't clear to us that they would. But it turned out to be— the system seemed to work beautifully. So we then turned to making antibody expression libraries on phage. And the library was a so-called random combinatorial library. It was a type that had been described earlier by the Scripps group, in which a library of antibody heavy chains is combined directly with a library of light chains, usually from the same immunization. And this process— so this process can be expected to generate the original heavy-light chain combinations of the B cell that it originated from, as well as entirely new combinations. And this is shown on this slide. So let's take a very simple case of two B cells. Encoding two different antibodies, one with a red heavy chain and a white light chain, and the other one with a green heavy chain and a black light chain. So if we clone out those genes and we shuffle them all up, we end up getting four possible combinations: two of the original and two new ones, the red-blacks and the green-whites. Now, as you increase the number of B cells that you use, you decrease the probability of finding the original combination, and you increase the probability of finding new combinations. So for example, if you took 1,000 B cells, in principle, which were all different, in principle, you'd get 1 million different combinations by combining them, of which 1,000 would be the originals, and 999,000 would be would be new combos. So we were very interested in new combos, because actually, if we were to take antibody genes from humans, we knew we needed to generate new combinations of heavy and light to generate new binding activities in order to be able to find things in there that would bind to the self-antigens. We knew the immune system would have deleted all of those original combinations that bound to the self-antigen. So for our first library, we took the V genes from the spleen of a mouse immunized with a hapten, phenyl oxazolone. Now, you may wonder what a hapten is. It's essentially just a chemical entity that's coupled to a protein. And it's been used by immunologists for many years to study how the immune system works. And it simplifies the understanding of of how the immune system works and the kind of antibodies that you get to different targets. Now, from a library of about 1 million clones, we obtained phage binders with a range of different heavy and light chain pairs. And we couldn't distinguish the original from the new ones, but some had binding affinities comparable to the very best hybridomas. In other words, in the order of 10 nanomolar. These were hybridomas made by César Milstein from that same spleen, because we ran the thing in parallel. We cut the spleen in half. He did one bit, we did the other. The affinities he could get were as good as the ones that we could get, which was a great irritation to him. But the technology was looking very promising. With our second library, we faced up to the altogether more difficult problem of making human monoclonal antibodies without immunization. After actually quite a few technical improvements I don't have time to go into, we created a larger and more diverse library, this time from the white cells of human blood donors. From a library of 10 million phage clones— so this is 10-fold bigger than the one I've described earlier from the immune mouse— we managed to isolate binders to many different antigens, foreign antigens, human self-antigens, but these had binding affinities in the micromolar range. And those affinities are too poor to use for therapeutic use. So we knew we'd got to improve those affinities. So I'm just going to focus on the left-hand side of the slide. We decided to improve those affinities. Let's follow what the immune system does. So the immune system takes selected B-cell clones, and it subjects them to random mutation and further rounds of selection with antigen. So for example, we took a phenyloxacillin-binding phage antibody from our human library with a rather modest micromolar affinity, and we grew it in a bacterial mutator strain. And after multiple rounds of growth and selection, increasingly stringent selection, we managed to get mutants crawling out at the end with affinities 100-fold greater, now in the low nanomolar range, which is quite comparable to the affinities of mouse monoclonal antibodies made by repeated immunizations of the same antigen in a mouse. We could even construct a genealogical tree, and we've got some attempt to show this here, in which you could see the sequences that we— by sequencing all the mutants, that we got at different rounds, we could identify 4 sequentially acquired mutations, each of which was responsible for lifting up the affinity and giving us this great increase in binding affinity. The right-hand side of the slide describes a shuffling approach, which I'm not going to go into, but it's an alternative and perhaps simpler way of creating— faster way of creating affinity improvements.. But it's not enough, or it wasn't enough, just simply to create mutations. We also needed to find ways of selecting between those phages with different binding affinities. And to do this, we applied increasingly stringent affinity selection. So for example, using longer wash times, so all the weak binders washed off, and the strong binders stayed stuck to the support. The other possibility was to use low concentrations of antigen. So here, what I've shown here, is on the left, is we've biotinylated the antigen, let it equilibrate with the phage, and then captured the phage bound to the antigen using streptavidin beads. And that also showed that we could— this, we could apply that and fish out selectively those phages with the higher binding activity. But there was a confounding issue, which is touched on here, and that is, that we discovered— we hadn't really expected this— whereas in George's case, the phage generally, all the antibody heads express a peptide. But in the case of the antibodies, we found that a lot of our antibody heads were chewed off. So a phage antibody might have between 1 and 5 heads, each species actually presenting different binding avidities to a a solid support, which greatly complicated our selection of those with the highest binding affinity. So we reduced this problem by an approach devised by Jim Wells, in which what he did was— I think he'd done it for a slightly different purpose— in which we created what's so-called monomeric phage, with fewer antibody heads. So actually what we end up doing is infecting— oh, sorry. Encode the antibody fusion on a so-called phage mid, and we use a helper phage to supply the other gene 3 subunits. So we end up with antibodies which have a single antibody on there, or otherwise they're completely bald. Now the bald ones don't bind to anything, so that's very easy. So you just end up with those which are bearing the antibody. So, As well as developing libraries of V genes from these B lymphocytes, we also had been exploring the creation of synthetic V gene libraries by a combination of the underlying V gene segment building blocks. Now, the advantage of synthetic antibodies is that the diversity of the library is known, and we can tune it to order. But at that time, The sequences of most of the human V genes were not known. And if we wanted to use this as building blocks, we had to find them. So we started mapping and systematically cloning and sequencing all of the human antibody building blocks. And then we assembled them to create our largest and most diverse phage library of more than 10 billion clones, including almost all of the human V gene segments. So we've now got a huge library, and we could select that. So we assembled the library, these V gene segments, as indicated on the left. And then we selected that library against foreign and self-antigens. And we now found antibodies with binding affinities in the nanomolar range, quite comparable to the affinities of hybridomas. In fact, really, by using these very large libraries, we avoided the need to do so much affinity maturation. So the great thing about this is, from a single very large library, you could fish out antibodies to all kinds of different targets. Now, in parallel with our work at the MRC, the startup company Cambridge Antibody was working with commercial partners to develop human therapeutic antibodies. Their most spectacular success was the development of the phage antibody adalimumab, which is otherwise known as Humira, which is targeted against TNF-alpha, an inflammatory mediator, and it's used for treatment of autoimmune inflammatory diseases such as rheumatoid arthritis, psoriasis, or Crohn's disease. Now, when we started the work on adalimumab, the random combinatorial light— the random combinatorial libraries of human antibodies just weren't large or diverse enough. We couldn't find anything in there that would be useful. So we had to try a different strategy. And we decided we would try and bootstrap starting with a mouse antibody. So we have a mouse antibody, which I've marked in red. And we took this, for example, here, the mouse antibody, heavy chain variable region. We paired it with a library of human light chains. So you end up with these hybrids, half human, half mouse. We're selecting on antigen. And we did, in fact, find we got some binders by doing that. And now what we did is we combined that selected human chain here with its partner chains from a human library, and essentially selected again. So in a couple of steps, we'd managed to create a human antibody essentially from a mouse template. So I wouldn't call it a humanized antibody. There's no trace of mouse in there other than we have built this human antibody using this mouse template. So it's essentially rather like this whole grandfather's axe business, you know, the handle, you know, that my grandfather changed the head and my father changed the handle. And, you know, it's my grandfather's axe. Well, you know, it's completely different. The whole thing is completely different. Identity altogether. Now, other than adalimumab, Cambridge Antibody Technology, and later MedImmune, into which it was incorporated, focused on creating very large libraries, usually not synthetic, but from human donors, of about 100 billion clones. And they end up getting high-affinity antibodies from those. And I'm grateful to Jane Osborne of MedImmune, who used to work in Cambridge Antibody before they were taken over, for the information on this slide. So the upper portion shows the range of pharmaceutical target classes, including targets here which MedImmune has made phage antibodies to. And you could see a whole range of growth factors, chemokines, ion channels, receptors, GPCRs, cytokines, protease inhibitors, and peptides. And on the lower portion of this slide, down here, you can see those phage antibodies that are already on the market. So they've gone through the whole process of clinical trials and testing. And there are 60 more still in clinical trial, still moving through the system. And so, in conclusion, you can see that this technology of these antibody repertoires with the filamentous phage has actually proved an extremely powerful way of harnessing evolution and is delivering a great new wave of antibody medicines. With that, thank you.