In Their Words

They’re acting like distributed processors, and the output is then next-generation sequencing. So the scale of the approach comes down to the fact that we can do everything in batch. We can synthesize DNA in batch. We can transform those DNA into yeast cells in batch. We can take those libraries and we can mix them all together. We can isolate all of the fused cells that indicate what proteins interact altogether. There’s no need to isolate individual cells. We’re isolating populations of cells using bulk selection. And then we can use next-generation sequencing to read out in batch. So instead of having a single well that measures one interaction, we can have a single flask that measures a million interactions.

Podcast Transcript

David Brühlmann [00:00:42]:
Picture a perfectly designed antibody that looks flawless on screen. Perfect geometry, ideal energetics. Then, fails to bind when it’s finally tested at the bench. Why does that gap between simulation and reality keep swallowing promising molecules? David Younger has spent years chasing that question. As the founding CEO of A-Alpha Bio and co-inventor of the AlphaSeq platform, he has built tools to measure protein interactions at massive scale. Today, we explore the real bottleneck behind AI-enabled protein engineering. Welcome, David. It’s good to have you on today.

David Younger [00:02:42]:
Thank you so much, David. Very happy to be here.

David Brühlmann [00:02:44]:
To start us out, David, share something that you believe about biotherapeutic development that most people disagree with.

David Younger [00:02:53]:
I think two things that I’ll share. One is that I believe that the amount of data that we need to solve really difficult problems, especially leveraging AI, is vastly underestimated by the field. I think that there is this idea that we can leverage historical data, that that’s going to be enough, or that maybe we need some relatively small but highly curated datasets to solve some of these really big problems in AI. And I think that what we as a company believe, after training models on different amounts of data and seeing how models improve, is that really to solve some of these really challenging problems, we’re going to need much more data than the field expects. So that’s first.

Second, somewhat related, is about data quality. I think that there is an expectation that data in the public domain, data from papers, is going to be relatively easily leveraged to solve some of these problems. And what many other people have found who have attempted to really leverage those public data is that oftentimes they’re not interoperable, or experiments aren’t done with the proper controls or the proper annotations. And so the quality of those publicly available data, with some exceptions, I think the PDB is an incredible exception, but for most public data, especially affinity data, the quality of those public data are just not as high as people expect.

David Brühlmann [00:04:14]:
Thanks for shedding light on that, and I’m looking forward to diving further into this topic. But let’s start with you, David. What first drew you into biotech, and what were some interesting pit stops along your career leading now to your role as a CEO of a company?

David Younger [00:04:30]:
I was inspired from a very early age. My parents are both in the medical field. My dad is a doctor, my mom is a nurse. And so I think part of this was early childhood dinner table conversations about medicine and patients that I found fascinating. Once I started studying biology, I went to Rice University in Houston and studied bioengineering. I realized that I was less interested in thinking about medicine at a bedside and more interested in how can we develop technologies, therapeutics that can have a much broader impact.

And then later on, as an undergraduate, I got exposed to the fields of synthetic biology and computational protein design, and just found these spaces just absolutely fascinating. The idea from a synthetic biology perspective, the idea that you could re-engineer an organism at a DNA level to get it to do something new and useful was just really mind-blowing to me and just got my creative juices flowing about how we could leverage this type of technology, especially as we get better and better at DNA synthesis and gene editing and understanding the relationship between various genetic mutations and organism function. And then computational protein design, somewhat similarly, the idea that you could engineer a protein entirely in silico and predict how it folds and predict its function, it really changes the way I thought about biology and how we could harness biology for a variety of applications related to human health.

So after undergrad, moved to Seattle and joined the Institute for Protein Design, David Baker’s lab at the University of Washington, was also co-advised by Eric Klavins, who runs the Center for Synthetic Biology at the University of Washington, and really was fascinated right at this intersection of synthetic biology, re-engineering cells, and applying that technology for applications that related to computational protein design. And what I observed as a graduate student kind of early on in the Baker Lab at the Institute for Protein Design was that here you had graduate students, you had postdocs, you had undergraduates who would regularly design hundreds of thousands, millions of computational proteins. And would then be downselecting to 12 or 24 to actually test in the lab.

And kind of recognizing how limiting that data bottleneck was for progressing research, mine started to go to how can we potentially think about synthetic biology as a tool for solving this problem? How can we use cells as essentially distributed processors as a tool for measuring, characterizing interactions between proteins in massive throughput that would really relieve this bottleneck for computational protein design? At this time, this was 2012. This was well before AlphaFold and kind of thinking about computational protein design with AI. It was really physics-based, first-principles computational tools. And so that need, the importance of generating those types of high-throughput datasets has only expanded enormously as AI had more and more impact in this space.

David Brühlmann [00:07:31]:
Let’s talk about the problem you alluded to. What I understand that still nowadays, a lot of companies start out with a lot of different potential molecule structures and then eliminate most of them, and they have a handful of candidates, and then finally they select their lead candidate. So paint us a picture of how today most of the people do drug discovery and I’d say protein engineering, what works well and what are the bottlenecks and why now we need a better tool?

David Younger [00:08:02]:
Still, I mean, most therapeutic proteins, most antibodies are discovered with conventional technology. So either an animal immunization, some very complex humanized animals, but still an animal immunization, or kind of a classic selection with phage panning or yeast surface display and yeast enrichment. These are still kind of the workhorses for discovering novel therapeutic antibodies. And it’s entirely serendipity-based.

We’re not going at this problem and saying, we’re looking for this exact protein sequence that binds this epitope at this angle with these exact properties. We are essentially saying, we know that we need something to bind to this antigen. We’re going to throw it into an animal, or we’re going to throw it into a screen. We’re going to see what binds. We’re going to downselect to a very small number of potential binders that come out of that screen and further characterize those and pick the one that happens to be best. If none meet our requirements, we’ll start from the beginning. And I think that the exciting thing about computational protein design and de novo design for antibodies is that we’re no longer having to be so dependent on serendipity. We can now say, we want a binder that targets this epitope in this particular way. We want this part of the paratope engaged. We want to also hopefully engineer these other properties into that binder.

I think that for data bottlenecks, there’s a data bottleneck on both sides. So for computational protein design, for de novo design, for in silico optimization, what we call zero-shot optimization, we need a lot more training data to get ML models to a certain level of performance such that they perform very well at these tasks. Despite the buzz and the hype, I mean, we’ve made a huge amount of progress, but we’re not there. Right now, we’ll see reasonable single-digit percent success rates for de novo designs against easiest targets and easiest epitopes. That is not the ambition of the field.

We want to be able to engineer against challenging epitopes and also engineer other properties. Success rates for those types of problems are still very, very low. So we need more data, we need better models. And then there’s also a data bottleneck on the other end, where if there’s a need to downselect to a very small number of candidates for thorough characterization, it means that we’re having to throw away a huge number of potentially interesting molecules that we just don’t have the resources to test using low-throughput experimental techniques. So we both want to generate much greater quantity of training data so that we can build better models and also have tools such that we can test designs by the hundreds of thousands or more, rather than having to downselect to a very small number of candidates.

David Brühlmann [00:10:44]:
So what is the solution now you’re proposing to solve that bottleneck?

David Younger [00:10:50]:
So we have a proprietary technology that’s called AlphaSeq that allows us to measure protein interactions at a massive scale. So this leverages synthetic biology. So we’ve re-engineered yeast cells, specifically re-engineered yeast mating. So yeast naturally exist as MATa and MATα haploid strains, and these 2 strains naturally bind to each other via protein interaction on the cell surface and then fuse. We have chopped those proteins off of the cell surface so that they’re no longer able to bind and fuse, and then we can replace with the surface display any arbitrary proteins of interest. So for example, a library of antibodies and a library of antigens. And now when we mix those 2 haploid populations, mating, so cell binding and fusion, is driven by protein interactions that we have engineered into the system.

And in so doing, we can take, for example, 1,000 antibodies and 1,000 antigens and quantitatively measure the affinities of all of the pairwise million interactions between those 2 libraries. So it’s a very, very powerful tool for quantitatively measuring protein affinity at a really unprecedented scale. That’s one innovation. And I think that when we talk about the bottleneck at the end of the process, right, of being able to characterize the order of thousands or hundreds of thousands of candidates, this is a very, very powerful tool for being able to measure at that scale. For training data, it’s also valuable, right?

We can generate an abundance of affinity data that can be used to train models. We’ve also figured out how to turn those data into structural data that we can talk more about, which is particularly important for training de novo design models. But even at the scale of measuring a million affinities in a single experiment, again, we believe that the data that’s really going to be required to solve these problems is far larger than people expect. So even at the scale of being able to measure a million at a time, we still think that this is going to be a challenging problem that requires some really creative solutions. And so in addition to technological innovation, we have also been thinking very critically about how we can leverage business model innovation to incentivize the accumulation of vast datasets using high-throughput technologies. But really incentivize the accumulation of those vast datasets to solve this problem, not just for a single pharma company, but hopefully for the broader industry.

David Brühlmann [00:13:06]:
That’s great that you’re trying to solve the puzzle from different ends. And we’re going to dive into the business aspects a bit later in our conversation. I would just like to follow up on how you’re solving, I’d say, the scientific, more technical bottleneck. What I try to do, I try to vulgarize a bit what you’re doing just to make sure that I fully understand. So instead of just finding a lot of data and then training your machine learning models, you actually found a scientific way to quote unquote automatically generate a lot of data from protein-protein interactions, which then you can feed into— so in a later stage to feed your model. But I’d say your approach is not only based on the modeling, it’s also some kind of experimental approach, right, at the same time.

David Younger [00:13:53]:
That’s right. At our core at A-Alpha, we are a data company. We generate huge amounts of data in the lab that centres on our AlphaSeq platform for measuring protein-protein affinities. And this is a technology that allows us to not just measure a single protein-protein interaction affinity at a time, as is done traditionally using biophysical techniques like bioluminescence spectrometry or surface plasmon resonance, where you express individual proteins, you purify those proteins, you measure binding one at a time. With AlphaSeq, we’re able to leverage synthetic biology to express proteins on the surface of cells, and then essentially have those cells act like distributed processors. They’re all in a liquid culture together. They’re all floating around and colliding with each other. And every time that they collide, they’re essentially sampling a protein interaction. If a protein interaction is strong enough, then those 2 cells will bind to each other and fuse. And we can use next-generation sequencing, DNA sequencing, to essentially identify all of those fused cells and measure affinities millions at a time instead of one at a time.

David Brühlmann [00:15:02]:
And how do you solve that analytical bottleneck? Because traditional methods are time-consuming. They’re expensive, especially if you want to measure that many interactions. How is your approach different in a much cheaper way?

David Younger [00:15:17]:
The traditional approach requires proteins to be individually expressed and purified and measured one at a time. So essentially you can think about in a single well, you can measure one interaction. With our approach, we’re leveraging synthetic biology. We’re re-engineering yeast cells to essentially do this characterization for us, where the cells are expressing proteins and they’re floating around in culture, they’re colliding with each other, they’re sampling interactions between proteins.

They’re acting like distributed processors. And the output is then next-generation sequencing. So the scale of the approach comes down to the fact that we can do everything in batch. We can synthesize DNA in batch. We can transform those DNA into yeast cells in batch. We can take those libraries and we can mix them all together. We can isolate all of the fused cells that indicate what proteins interact altogether. There’s no need to isolate individual cells. We’re isolating populations of cells using bulk selection.

And then we can use next-gen sequencing to read out in batch. So instead of having single well that measures one interaction, we can have a single flask that measures a million interactions. So far more scalable, far faster on a per-protein-protein interaction basis, but we’re still able to capture highly quantitative affinity data from this technique.

David Brühlmann [00:16:41]:
And if we do a head-to-head comparison between the traditional way of protein engineering and using your approach, how does this translate into a greater speed or cost, or what are the main differences?

David Younger [00:16:54]:
So specifically for measuring protein interactions, it’s about cost per protein per affinity measurement and scale. So where you might otherwise be measuring one protein interaction and with kind of all involved of engineering a cell line, expressing those proteins, purifying those proteins, performing SPR or BLI, you’re talking about hundreds of dollars to measure a single interaction. Here we’re talking about measuring a million interactions in a single experiment at a far, far more efficient per protein-protein interaction cost. So this allows us to accumulate enormous amounts of affinity data. And we can also leverage these same types of data to get structural data as well that are really important for some machine learning models. And with this technique, with the ability to measure a million affinities at a time instead of just one at a time.

We can really rethink how we generate training data. We can rethink how we need to downselect potential therapeutic candidates coming out of a pipeline to find the ones with ideal properties. We no longer need to downselect to 24. We can downselect to 100,000. So I think with this type of a capability, we can really rethink how we approach both the generation of training data, but also therapeutic discovery and optimization.

David Brühlmann [00:18:15]:
How are you going to train these, I’d say, generalizable de novo design models with the much greater availability of data?

David Younger [00:18:23]:
So for computational protein engineering, we have really focused on 2 primary problem statements. So these are de novo design. So de novo design being: we have a target, we’ve identified an epitope on that target. And we want to fully in silico engineer a binder that will engage that epitope on that target. So that’s de novo design, essentially going from 0 to 1. And then the second problem is zero-shot optimization. So here we have a binder that engages a target, and our goal is to, without any problem-specific, project-specific data, predict and implement mutations to that binder that will improve affinity and improve other properties.

And these are two problem statements that are very closely related. You can imagine in a kind of full de novo design pipeline, we might want to first have in silico de novo discovery to find binders to an epitope. And then before we ever move into the lab, move those most promising ones into zero-shot optimization to try to come up with an even stronger binder than we otherwise would have. So closely related problem statements, but we believe that the kind of fundamental underlying data that’s needed to solve those 2 problems and the kind of ways to think about models for those 2 problems are actually quite different. For de novo design, what we have seen and what others have seen is that structural data is immensely valuable. So de novo design models are traditionally trained from PDB data. The problem is that PDB data, while incredible in quality is not so good in abundance or diversity. So for the PDB, we’re talking about today about 10,000 antibody-antigen structures in the PDB, and that’s growing at about 1,000 structures per year.

David Brühlmann [00:20:14]:
PDB, what does it stand for?

David Younger [00:20:16]:
Protein database. So this is the database that protein engineers and biochemists will turn to understand structure of proteins. And this is a really incredible resource. So it is an accumulation of crystal structures that have been solved over the last many decades that have all been accumulated into this single really incredible resource that has been central for really all of the advances that we’ve seen in computational protein design.

Even so, it’s very limited. So 10,000 antibody antigen structures, about 1,000 new depositions per year. Some of the work that we’ve done internally to show kind of scaling laws of how de novo design models improve as we add more and more data seem to indicate that for really good generalizable de novo design model performance, we’re probably going to require on the order of hundreds of thousands to millions of structures at a baseline. So if we’re talking about 10,000 today and 1,000 added every year, this is obviously far, far too long to wait to get to the type of model performance that we’re hoping to achieve.

And so we and others have really been thinking very hard about how can we use higher throughput wet lab technologies to generate the types of structural training data that’s going to be needed to address that problem. So de novo design is one of those problems, and structural data, we believe, is the key need. The other problem is zero-shot optimization. The PDB, while it contains structural information, it contains essentially no information about how strongly those proteins interact. So you can see physically how they interact, but you don’t know how strongly.

It also doesn’t contain any local changes to those proteins, local mutational landscapes. It essentially has completely distinct parental antibodies and their antigens. I don’t think that there are any examples, maybe there are a very small number with some COVID examples, where you have an antibody binding to an antigen and you also have the structure of an antibody binding to an antigen that’s only 1 or 2 mutations away from that other, to see how the structure changes, for example, when you’re making only very, very subtle changes to those proteins. And again, even if there are some examples from COVID there aren’t affinities associated with those structures.

And so what we believe is needed for zero-shot optimization is a real quantitative understanding of how mutations impact bindings. And so we can generate datasets like taking different parental antibodies, making all single-point mutations to those antibodies, and then measuring how all of those different variants impact binding to a given target, and then repeat that over and over and over again for many, many targets.

And it’s these types of accumulated, what we call local affinity landscapes, that we have seen really start to move the needle in model performance for zero-shot optimization.

David Brühlmann [00:23:00]:
There’s one more aspect I’d like to touch upon regarding models. A lot of people talk about foundational models. How do you know if one model is better than the other?

David Younger [00:23:10]:
It’s a huge problem. I think we are right now as an industry with a number of different companies and academic groups, all approaching the same problem of de novo design and AI protein engineering. Each of them have their own techniques, but each of them also have their own disparate evaluations. So when you read papers where a company will talk about their performance and how they are now the leader in the field, it’s typically an experiment against very, very defined, very unique bespoke set of targets.

And the way that they are evaluating success is not at all comparable to other companies. And so it’s very, very difficult. There’s a huge amount of confusion to sort of actually know which are the models that are best. And not necessarily is there going to be a single model that’s best at everything, right? There might be some models that are best at a certain type of target or certain type of epitope. And there might be other models that are better for a different class of epitope. There might be some models that allow you to get stronger starting points. Maybe the hit rate against the novel epitope is lower, but they’re able to produce higher potency binders because they have kind of more sophisticated zero-shot optimization embedded in that model.

Whereas others might have higher hit rates, but you’re all looking at very, very weak binders that are coming out of those pipelines. There is currently no standard, no universal benchmark for evaluating performance. And this leads to a huge amount of confusion across the field. I think pharma companies who are considering working with one of these model makers are certainly confused about which company they should be working with. The model makers themselves are confused, to be perfectly honest, about how does our model actually stack up against others in the field? So it’s a really a big problem. And we believe that universal industry-wide benchmarking is going to be essential in order to move the field forward. And this is another thing that we are very actively engaged în. Currently, we’re working with a handful of companies on a benchmarking study where our goal is to standardize the targets and the difficulty of targets that we’re evaluating models against, and then really kind of providing true apples-to-apples comparisons.

The other challenge with benchmarking is that if companies are only able to characterize at low throughput, if they’re only characterizing 96 proteins coming out of their pipeline, the error bars are just absolutely massive. It doesn’t really give us a deep picture into how those models are performing from a true statistical perspective. Whereas in our benchmark, when we’re working with some of these companies, we will be able to characterize tens of thousands of designs coming out of their models and hopefully get a much clearer picture of how these various models are performing.

David Brühlmann [00:25:55]:
The gap between a beautiful in silico design and the working molecule often comes down to one thing: not enough real interaction data to learn from. David Younger’s work is a reminder that fixing AI-enabled protein engineering takes better science and substantially more data. If this resonated, leave a review on Apple Podcasts or your favorite platform. Thank you so much for tuning in today, and I’ll see you in part two.