Merck & Co., Inc. (MRK) Earnings Call Transcript & Summary

Hello. This is -- I'm Bernhard Geierstanger. Welcome to the live panel discussion on protein engineering and molecular design. I have with me Jamie Spangler from Johns Hopkins University; Brandon DeKosky from the University of Kansas; and Tim Whitehead from the University of Colorado in Boulder. I'm Bernhard Geierstanger, I'm with Merck Research Laboratories. I've been at Merck for only a little over a year. Before that, I spent 20 years at the Genomics Institute of the Novartis Research Foundation. So I was part of Novartis Research in San Diego. My background is that of a biophysical chemist, but I got more and more into protein modification technologies. I worked a lot of Unnatural Amino Acid Incorporation for various aspects, protein conjugation. Got very much involved in antibody drug conjugates, site-specific conjugates. And later on, the development of biotherapeutics, which included ADCs, antibodies, bispecifics, fusion proteins, pegylated proteins, enzymes, all of the above. At Merck, I worked -- I lead a group called Protein Engineering and Modalities. And the purpose of the group is to be modality agnostic and enable Merck in different modalities. So ADCs, again, is a big, big part of my efforts. As we develop molecules, one thing that we really stress is diversity of molecules. Just have a broad set of molecules in a drug discovery project with different properties, different sequences, and then being modality agnostic to the indication. So whatever the biology requires, we'll build. At the end of the day, we need to manufacture molecules. So for us, weeding out the bad apples in the beginning is a key aspect of it. And so emphasizing good biophysical properties, even in candidate molecules as we start to test them, is key for all of us at Merck, and I think, throughout the industry. And so we want to have a lively discussion here. So I hope we're going to get some good questions from the audience. And then all of us have different approaches and different backgrounds, so I hope we can have a lively discussion from all different aspects of protein engineering. And so with that, Jamie, perhaps you could introduce yourself and give a little bit of a perspective on your work and background.

Thanks so much, Bernhard. So excited to be here today to be part of this panel. So my name is Jamie Spangler. I'm an Assistant Professor of Biomedical Engineering and Chemical and Biomolecular Engineering at Johns Hopkins University. And my training is actually in both biomedical engineering and some coloring of chemical engineering as well. And that's really sort of established what my lab is about, which is sort of engineering and designing new molecules that do better than what's in nature. And we use a variety of different approaches. We do a lot of direct evolution-based work, particularly using the yeast surface display platform to discover and design and engineer new molecules. And we also combine that with structural insights that come from various -- either experimental structural methods such as X-ray photography, NMR, cryo-EM. And then also with computational methods, protein structure design prediction through collaborations with more computational groups. And then really seeing how, to Bernhard's point of sort of not just kind of doing that in a vacuum and then hoping that something will happen downstream, but to really thinking about how we get molecules from the discovery through the design and implementation, all the way through to the translation, and ultimately, to the clinic. So that's just kind of a general background of my training and happy to address any questions you guys may have.

Tim, do you want to go next?

Yes, absolutely. So my name is Tim Whitehead, and I'm an Associate Professor at the University of Colorado in Boulder. And my training was as a chemical engineer for Bachelor and PhD. And then I moved to work in the field of protein design, computational protein design biochemistry in the lab of David Baker before I began my independent career. And so my current research interests are really twofold. We really apply the tools of protein engineering design to understand how antibodies recognize and respond to viral pathogens. And we look at typically the antigenic proteins on surfaces of enveloped viruses. The other portion of my lab, which may not be of this interest to this -- to the participants on this panel, we really focus on chemical genetics where we try to redesign proteins to respond and activate in response to small molecule stimulus where we can control this small molecule that can control protein activity. And so we're using this from everything from redesigning how plants sense and respond to xenobiotics to actually making small molecule responsive bacteriophage for targeted-based therapy.

Brandon?

Great. Yes, and thanks again to all the participants for logging in. Excited to have some great discussions today, and it's great to be on a panel with Jamie, Tim and Bernhard. So our group -- my background is in chemical engineering and biochemical engineering, a little bit of work in industrial bioprocessing. And then I switched to single cell analysis. And so we got very heavily involved in a single cell analysis of paired heavy and light chains from antibody libraries. A network was done at George Georgiou's group and UT Austin during my PhD. And then from there, we transitioned that to functional screening. So taking those paired heavy and light chain libraries from human patients for animal donors and performing actually functional screening to map at a large scale the different functions of antibodies in an immune response. And so we've been working very heavily on that. We've been screening against various viral interactions and autoimmune interactions. And more recently, we've also been looking into kind of understanding the pathways of antibody development in vivo, looking at aspects of somatic hypermutation pathways and trade-offs and fundamental mechanisms that may guide antibodies to mutate in one particular direction or another. And also looking at things like personalization of antibody immune responses. So why is one person's antibody response different from somebody else's? And how can that help us understand the basics of immune mechanisms and eventually understand the -- how to design and improve personalized immune therapeutics? So I think that's a direction that we're definitely heading. So thanks again for the panel and look forward to all these questions.

You see great risk background actually in terms of how we all approach this. I heard computational designs of proteins a couple of times. So I don't know, Tim or Jamie, perhaps you want to give us a perspective on where you see this going? How mature is the field in terms of designing proteins or binders by de novo operation calculations and computational approaches?

Jamie, I think that's been for you first.

Yes. No. I mean, honestly, as an experimentalist who's computational adjacent and having the opportunity to work with top protein engineering design labs, I've just been astonished honestly in how much progress has been made even over the last like 5 years, much everything you see coming out of [indiscernible] is de novo switcher, de novo [indiscernible] de novo sort of being designed. And what's really exciting and what's really been very pleasantly surprising, I think, in my mind, is how accurate a lot of the computational predictions are for the actual structures of these molecules coming out. And I think particularly when we think about cytokines for bundle sort of molecules, it's remarkable how much agreement you have between the computationally predictive structure and then what actually happened. I would say also, I'm really -- I would say like there's a huge element of synergy between the computational methods and then the directed evolution methods that sort of carry it to the next level. Like what the computer spits out maybe 20, 30 different sort of top molecules. But then there's definitely multiple grounds of sort of maturation that need to happen to really get it sort of up to stuff and also to make sure that it meets the design requirements. So it sort of kind of gets you in the ballpark. But then to really hone in more, it's important to use the directed evolution tools that you had available to you as well. So I think there have been numerous examples of projects where you sort of get something different or unexpected from the computational angle, but then in order to make it feasible or practical requires some tuning up. And that's actually also related to some of Bernhard's interest about manufacturability and also stability of these different molecules. So they're designed to be hyper stable. But then sometimes, what happens is when you do the directed evolution on them, they actually become less stable. They actually destabilize the protein. And so there are certain approaches that people have more recently within the directed evolution process to try to engineer, not just for binding and function, but also for stability. So I think all of those are important things, but interested to hear sort of Tim's perspective on some of the computational tools, having worked firsthand with some of the design tools and also with Brandon's sort of perspective on sequencing and how sequence prediction can come into play as well.

So what would you say is kind of the upper limit of size? What are the restrictions on the molecules that we can actually successfully work on?

I don't know if it's so much size, so much as like kind of secondary structure and kind of the types of domains that we're trying to predict. I don't think there's like -- sort of like a maximum. You make an infinitely long like some sort of regular impeded structure. But I think there's a lot more difficulty with designing interfaces and loops in other regions that might not be as predictable. So yes, I need to not be able to put a direct number on it. But I think it's a little bit hard without knowing the context of how flexible or how rigid the particular -- in order the particular range would be.

Yes. I mean what I would say is that size is really not an issue. It's loops and it's polarity, right? If you moving the backbone and changing the backbone in doing this with a -- doing this with -- for something that matters like antibodies is a very, very difficult problem. And then also dealing with polarity. And so if you have a discerning eye and you look at kind of the solutions basis of things that have worked in the computation design field, most of the things that work are things like interfaces where you're binding with very hydrophobic interactions, right? And so that's a big challenge. What I would say was really exciting is the ability to now sequence entire antibody repertoire. And so the big advance computationally in the last year or so is a Google team called AlphaFold has really, in my mind, nearly solved the protein volume problem, which is that if you have an existing sequence of amino acids, can you determine the 3-dimensional structure for an existing protein. And they've done that in large part by deep learning, using both structural data as well as sequence data. So massively over sequencing environmental samples. And so you can do the same thing for these antibody repertoires like Brandon is doing. And if you can do that, you can actually harness information for prediction and for design. And I think that's really where the field has been moving in the last couple of years, and I expect to see some really fantastic papers coming out really harnessing that power for prediction of antibody developability as well as binding.

And the one thing that I might add on to a lot of that is that the antibody binding protein-protein docking problem, even though there's a lot of structure, there's still a lot of unknowns in those CDR3 loops. And at the somatic hypermutation, it can compound it a little bit. And I think in terms of the forefront of maybe like personalized drug discovery and personalized drug design using computation, maybe in the T cell receptor space, where there's been some incredible progress because of the amount of templated interactions that are there already between the TCR and the MHC2 and the known prevalence of different peptides finding. And so in some ways, I see the antibody computational docking problems following TCRs a little bit, it's a little bit lower risk because you can just get rid of all of the bad designs at the lab scale. And so I'm just really excited to see what happens with computational development of like personalized TCR drugs and things like that or even analysis in the next few years.

So do you think T cell receptor interactions are easier or harder? I mean, from -- it seems to me that getting specificity may be actually pretty hard for that one.

Yes. So actually, that's a very good point that you bring up. TCRs have a much higher complication from antibodies in that the affinity of interactions is often much lower, and that's a complication. And then also, yes, because you've got that MHC backbone that has that conserved interaction across MHCs that are with other -- the same MHC bound to a different peptide and also peptide similarity, that can also get a little bit tricky. Yes, a lot of it is driven by that affinity difference. So certainly, there are some complications there as well, yes.

So from -- I stressed in my opening statement that we're looking for molecules by physical behavior. And so their even point mutations or introduced [indiscernible] is a big issue. So you want a molecule that don't tend to aggregate form high molecular weight species, et cetera. You want molecules that are more stable as a surrogate for aggregation in good pK. Is -- are the computational tools, do they tend to over predict hydrophobic interactions? Or is it the other way that the more specific interactions like hydrogen bonding or something dominates? Is there any data out there that gives you guidance on how to use the predictions?

With respect to developability, is that specific...

Yes. Yes, so I mean, has anyone looked and actually tried to -- are we getting binders, for example, that are competitive with things that have evolved in a mouse?

Yes, so there are a number of sequence-based predictions on things like aggregation propensity and things like solubility. So they're measuring biophysical terms like a second [indiscernible] and so on. I'm not an expert in those areas. So people like Pete Tessier at University of Michigan or [indiscernible] at Cambridge and others are kind of at the forefront of those fields. I'm not sure how fantastic this perform. I think one thing that we're doing right now, which is -- which I'm excited about and we haven't published on, is we've been looking at the repertoire data from humans and we've just asked, with a nod to developability, whether mutations in framework regions for clinically developed antibodies have those framework mutations are overrepresented in the repertoire data. And the idea is as follows: there are a lot of developability routes for regulatory approved antibodies. So mice -- humanized mice, display and many others. And during that development, a lot of times, people have made point mutations or changes in those framework regions to go and enhance developability. And the question we asked was just could you predict those mutations just from looking at sequencing repertoires themselves? And the answer was yes. The answer is that the things that actually make it to the clinic are overrepresented in the -- if you just pull those repertoires and look at those mutations in the framework regions. And so that's just a very easy way for developability, right? Just look at the repertoire data sets and just choose what has nature has already chosen for you, right, to do.

Brandon is smiling there.

All of these are just fantastic topics that are exciting to talk about. So yes, I guess, the one thing I would add on to that is with those biophysical determinants, there's a few different ones. Some of them are -- as Tim alluded to, some of the V genes that are less represented, that can be somewhat of a flag since they're known to be a risk for clinical development. There's things like surface -- excess surface positive or negative charges, I think positive charges, especially and excessive hydrophobicity in a number of others. And those folks that Tim mentioned have been leaders in trying to understand that. And each one is thought of like -- it's kind of like accumulation of different yellow flags. And if you have a molecule that has too many of those yellow flags, it's thought to be quite high risk. But any one of them on their own is not enough to conclusively say this antibody is developable or not. And it's just kind of like accumulation of risk, almost like if you think about sun exposure kind of accumulation, but you can't say yes or no. And then I see a question from [ Omer ] that I might field. The question is, I could not understand that if the sequencing of monoclonal B cells carry out, is it hard to determine the antibody sequence of the CDR? If it is hard to determine, what kind of approach can be done? So [ Omer ], that's a great question. We use NGS for sequencing single B cells routinely. And I think, as you mentioned with next-generation sequencing, there's a lot of sequence error that comes with those -- the next-generation sequencing technologies. And so it can be a little bit difficult to determine. For example, if you were to take a monoclonal antibody like a B cell line and sequence it a million times, you might have 100,000 different sequences that came out of the sequencer just from one cell. And so we use a lot of different error correction methods and quality filtering methods. And then you can get that number down substantially. And so quality filtering the data and then having a minimum read count and then maybe some sequence clustering, and that can get that number down to maybe a handful. And then if you were to generate a consensus sequence, you would, for sure, get back your original known control. So it has to do with like how many reads do you have. Because the more reads, the more error you're going to have. And what are your -- what are you looking for and things like that. But in general, if you use a consensus sequence, built in the right way, you will always return exactly what went in. So as long as you're using the power of numbers and statistics to rebuild the original sequence, you'll get it back 100% right every time.

Yes. Sorry, just jumping back on one of the points that Brandon made earlier about sort of the different flags and the different rules, yes Pete Tessier has done a lot of work in developing these different rules and actually using the data set from Adimab of 137 clinical stage antibodies and looking at are there particular rules and generalizations that we can draw. And so again, just to draw attention to that paper for people who are interested in thinking about that in more detail.

Though I'm pretty sure everyone who's listening to this talk has probably seen that paper, frankly, right?

Yes.

NAS 2017, right, something like that?

Yes, there were -- yes.

I'm going to have to make a pitch for a paper by my colleagues [ Mark Bailey ] and -- under the leadership of [indiscernible] of they worked done -- they did a similar study on 154 antibodies. And so I think it's in that 2020, so April of 2020. So I think it's -- will be interesting to compare the 2 studies, try to do some of that. One of the takeaways that I have from this type of study is basically standardization and a generation of those data sets, this high quality inconsistency. So setting up high throughput analytical assays that you just routinely run the same way for large sets of molecules so that you actually start building out data sets for AI or machine learning analysis, along those lines, so I think is a key aspect of it.

Yes. I know one of the limitations of a lot of these studies is that we have very good data for things that proceed through Phase III and get approved. But the earlier stages of the pipeline are a little bit more difficult because it's hard to tell, you can -- I believe the assumption is often made. These are the molecules that are in Phase I. And the assumption is that some number of them will fail, but we don't know which ones are going to fail and what the reasons are. And in some ways, it would be interesting to kind of close the loop, maybe there hasn't been enough time past. But in a couple of years, when we have more information about these molecules to see starting from that initial batch, these were the predictors and see if that holds true, if the characteristics of that set change as that particular subset goes in through and some of them get approved. So -- and I think that exactly the limitation that you had mentioned, Bernhard, as in terms of how do we get the characteristics of those that truly do not proceed in the clinic versus those that are delayed for various other reasons.

So this brings to mind my favorite saying about statistics, which is that if you go to the NBA, there's no correlation between players' ability and their height, right, which you would -- would lead you to say that there's no correlation between height and your ability to play in the NBA, which is not true, right? And so I think that's -- the key is trying to get access to -- for academics, certainly, I'll make a pitch to earlier stage in the pipeline to really be able to evaluate things that go and develop and things that don't. I think that would be really hugely valuable.

Yes. And also understanding the specific reasons why things aren't carried forward like sometimes it has to do with other post translational modifications, that sometimes gets kind of shoved under the rug and just kind of move on to another molecule. But yes, no, I think that's good to -- transparency is the only way we're going to make progress here. So that's great.

So I missed part of the discussion because I -- my interconnect -- Internet was unstable. Can you guys hear me right now?

Yes.

Okay. So for -- like if you had -- for us, going into like an antibody campaign, there's a diverse set of molecules upfront is key, or even like design type of applications, just because we don't really understand the rules and they make the molecular developable. And strategically, if it's hard to express the protein, you just pick the one that's, a, if you have choices, you go with the flow, right? You take the ones that make it easier as long as the function is still there. Obviously, the biological function is really the driving force initially. But then as the funnel gets narrower and you've zoomed in on tens of candidate molecules, then you pick the ones that don't have post translational modification issues or potential issues. You pick things that are less prone to aggregation, perhaps have a higher stability. So it's really risk management that you do at that point. You just -- among the options that you have, you pick the one that is less risky. I mean investment gets higher and higher as you go into making a -- master cell bank is very expensive. As you go further down development, it gets -- each step gets more evolved and more expensive. So it's actually really important to weed out the bad apples upfront and hopefully pick a really good molecule in the beginning.

Bernhard, actually, I have a question for you, it that's okay. But you having the most sort of development side experience. Are there certain flags or certain biophysical properties that are considered more important or less important than others? Because there's a lot of ways, right, that a protein can be ill behaved. But are there certain properties that might be -- drive that decision more than others?

Yes. Aggregation propensity is really high up there, I would say. And then also several unfolding, at least, above the cutoff that you're setting for a particular class of molecules. Fragmentation clipping is something that we look for early on, especially as engineered molecules. I think as soon as you have a few linkers or you have like a single chain in your construct and all that, I think you start worrying about those things. Personally, I'd like to stay as natural as possible. So I prefer natural folds over overly engineered fold spread.

Bernhard, do you see -- those rules have been figured out for monoclonal drugs on their own. Do you see any changes to them when you're looking at ADCs? Does that flip the script in some ways?

No. No, it really doesn't because, well, it adds complexity to it because you're putting a drug on to the antibody and the drug intrinsically hydrophobic character usually. So minimizing that impact of having a drug on top of something that is intrinsically relatively hydrophilic is key. So you can get -- you can do that by designing your linker and linker payloads appropriately or selecting more hydrophilic linker payloads, or you can get around it by using site-specific methods where you hide the drug. I personally like to put the drugs close to the surface of the antibody so minimizing the additional hydrophobic surfaces.

Do you run into a lot of issues targeting maybe hydrophobic sites with hydrophobic exposed loops with ADCs? Is that kind of something you have to really avoid?

I think most people, they pick a few preferred sites for site-specific attachment and it becomes -- you optimize around the confines and parameters of your platform.

I see there's some questions in the chat, I just noticed. Oliver Hill, thanks for the question. Says, what are the driving forces to separate folding units and secretary pathway-dependent proteins? And then a sequence order from N to C-terminus, what is your experience kind of be predicted? Interesting question.

I think that relates with the next question, which is the expression levels of design recombinant proteins can may be predicted. Because I think that's asking more fundamentally the second question. I would say that -- with the caveat that a lot goes into the expression levels for recombinant protein. So there's host definitive effects. There's medium effects. There's a number of different things, right? We're all aware that the biophysical property that sequence actually matters tremendously. And so I'll -- we've done some work in this area, but for things like antibodies where they're solved structures or other proteins with solved structures, kind of the best, the best tool out there is worked from my colleague, Sarel Fleishman at the Weizmann Institute. And his group has used computational design to predict mutations on given proteins that can enhance expression levels, soluble expression levels. This works the level of thermodynamics. So it improves the stability of the ground state or the folded state relative to the unfolded state, but it also works at the level of folding rate, which is really important for secretory pathway dependent proteins, right? Where once you are -- once you're translated fully, the rate at which you fold matters tremendously for success of secretion. And so the way that, that works -- so it's called PROS or protein repair one-stop shop. The dominant signal is evolutionary conservation. So if you look at mutations that -- for candidate mutations, there are ones that are predicted from the evolution of that protein sequence to be overrepresented and [indiscernible] in that given sequence. There are other things under the hood, but really that's a dominant signal. So that would be my suggestion.

We have one more. Any other comments on that expression? I think a lot of it has to do with initiation of translation. So I mean, people use different sequences upfront. So that has a huge and -- can make a big difference on top of all the things that Tim alluded to. There was -- there's one question about -- whether we can give feedback on aptamers and nanobodies as compared to antibodies. I assume this is around recognition. So nanobodies, I think, are very good finders. They're very compact. So there's -- they're quite attractive, in my mind. They're naturally evolved domains. You can get them very thermostable. Aptamers are more -- they're engineered like somalogic aptamers and so forth. They can recognize proteins quite well, too. So I'm not aware of -- I'm not sure that I can really compare that to you.

Maybe...

Go ahead, Jamie.

Oh, no, no, I was just going to say like maybe in terms of developability as I thought -- or if that's the question, I don't know.

I'm aware that aptamers have been used as detection reagents for like protein -- serum protein regimens and so forth. And for those applications, they are fine reagents.

Yes. I think, historically, aptamers have run into some difficulty in the clinic with their specificity. I think the binding affinity is not quite as high as antibodies, and their specificity is also not quite as high. And that's been a limiting factor historically. That's not to say that it prevents new applications, but I think that's where some of the trouble has been. And in terms of nanobodies, I think the farther you get from the native human monoclonal form, the more potentially you have for an immune response. And so I think nanobodies are -- you can run into that as an issue for repeated dosing of the molecule. But when it comes to like a single dose, I think nanobodies have some clear advantages in terms of penetration of tissue because the smaller size allows for getting to places that the larger antibody molecule can't get to. So that's my high-level impression of those 2 areas in the clinic.

Yes. Also in terms of nanobodies, there's some things in binding interfaces that nanobodies like particularly for GPCRs and other intrathecal proteins that nanobodies can cover that a full monoclonal really can't do the job. And so the way that I usually view it as sort of where there's a will, there's a way sort of things. So if there is a really effective nanobody that does something that an antibody cannot, I'm hopeful that we can find a way to translate that, even though it is in a different format that's less recognizable. Sorry, Tim's going to say something.

Great. I think we're out of time. So thank you. One thing that we didn't talk about is immunogenicity, which came up in the very last minute here. But I think that's a really long discussion. So anyway, thank you so much. Thank you to the audience for the good questions, and thank you for -- to you guys for a lively discussion. So take care.

Thanks for leading, Bernhard.

Enjoy the rest of the conference. Bye-bye.

Bye.

Thanks, everyone. It's a pleasure.

Bye.