🔬超越 AlphaFold：Boltz 如何将药物发现的未来开源

Actually, we only trained the big model once. That’s how much compute we had. We could only train it once, and while the model was training, we were finding bugs left and right.

A lot of them were bugs that I wrote. I remember us doing surgery in the middle: stopping the run, making the fix, and relaunching. We never actually went back to the start. We just kept training it with the bug fixes along the way.

Yeah.

Right.

Yeah.

It was impossible—

Yeah.

—to reproduce now.

Yeah, no, that model has gone through such a curriculum that it’s learned some weird stuff. But somehow, by some miracle, it worked out.

It’s a pleasure to have with us today Gabriele Corso and Jeremy Wohlwend. They recently founded Boltz, a company trying to democratize and bring structure prediction in biology to the masses. They’re both recent PhD graduates from MIT and have been working on all sorts of foundational papers in generative biology. Anyway, it’s a pleasure to have you here. Thanks for coming.

Thank you.

Thank you.

I guess we’re maybe 6 years post-AlphaFold 2 right now, which was kind of a big moment. Is that right?

I think it was 2021.

Yeah.

So, yeah, going on 5 years.

5 years.

Yeah.

5 years, yeah.

Yeah.

So maybe for the audience, let’s go back to that moment in time and explain what this big moment was and why it was interesting. Why was everyone so excited? I think you two were probably quite excited, so why were you personally excited?

1. Why AlphaFold Changed Biology

I’d start with why that was interesting from a scientific standpoint. Maybe first, as an introduction for those in the audience who are not structural biologists: the idea of structural biology is that we want to try to understand how proteins and other molecules take shape inside our cells and how they interact.

Structural biology is this beautiful discipline where we’re somehow able to understand these minuscule structures at atomic detail using incredibly complex methods like X-ray crystallography. The dream in computational biology has always been: can we understand the structures without having to resolve this crystal, shoot X-rays, and so on?

AlphaFold was a real breakthrough in this problem of protein folding, which is trying to understand the structure of a single protein. To me, it was exciting across many dimensions. First, I was a computer scientist working a lot on machine learning, and I saw the impact that work somewhat similar to what I was doing could have on a longstanding scientific problem.

Second, from a more personal perspective, seeing the structures coming out of these models—seeing this beautiful creation of life—was very inspiring to me. That was one of the things that led me to start working on structural biology, and in particular with machine learning.

Were you a structural biologist before AlphaFold came out? I mean, you did machine learning, but it wasn’t in structural biology, so that actually shifted your career quite dramatically.

Yeah, very dramatically. I was working on some pretty theoretical, methodological things, and I was starting to see some of the challenges in doing somewhat theoretical or methodological work and seeing the potential impact of applied machine learning.

AlphaFold was really a machine-learning breakthrough, but it was also applied machine learning, and so that led me to want to start working in applied ML.

Our group at the time was working a lot on small molecules already, and I think AlphaFold is what triggered this shift to working on biologics. At the time, I think it opened as many questions as it answered.

The immediate follow-ups were, “Okay, can we do this on things other than proteins? Can we model interactions of small molecules with proteins, nucleic acids with proteins? Can we model more complex protein systems?”

Very rapidly after AlphaFold, people realized that machine learning could really target this problem very differently than previous methodologies.

Going back to the AlphaFold 2 moment, I remember this very well. I was at NeurIPS when, I guess, the results of this famous competition came out. So I wanted you to talk about CASP, what it is, and why it was so interesting and exciting.

2. The CASP 14 Breakthrough

Every couple of years, the goal has always been to find protein structures that are a little bit different from what’s known. CASP, over the years, has put in a lot of effort to gather structures from academic groups and even industry groups to try to create a test set that would be difficult for different methods.

CASP 14 was when AlphaFold 2 really blew everything out of the water. The improvement was so large over the previous method and also over the previous competitions.

Now CASP continues. We’ve had CASP 15, and we have CASP 16. What’s happened now is that it’s really expanding to all these other modalities, as I was mentioning: proteins, small molecules, nucleic acids, and so on. But the goal remains to really challenge the models: how well do these models generalize? We’ve seen in some of the latest CASP competitions that, while we’re becoming really, really good at proteins, especially monomeric proteins, other modalities still remain pretty difficult. And that’s really essential in the field: that there are these efforts to gather benchmarks that are challenging, so it keeps us in line about what the models can do or not.

Yeah.

Yeah.

It’s interesting you say that. In some sense, at CASP 14, a problem was solved—pretty comprehensively, right? But at the same time, it was really only the beginning. Can you explain what the specific problem was that was arguably solved and what remains, which is probably quite open?

3. Structure Prediction Is Not Folding

I think we’ll steer away from the term “solved” because—

Solved.

—we have many friends in the community who get pretty upset at that word, and I think fairly so. But the problem on which a lot of progress was made was the ability to predict the structure of single-chain proteins.

Proteins can be composed of many chains, and single-chain proteins are just a single sequence of amino acids. One reason we’ve been able to make such progress is that we take a lot of hints from evolution. The way the models work is that they decode a lot of hints that come from evolutionary landscapes.

If you have some protein in an animal and you go find the similar protein across different organisms, you might find different mutations in them. As it turns out, if you take a lot of these sequences together and analyze them, you see that some positions in the sequence tend to evolve at the same time as other positions in the sequence—a correlation between different positions.

It turns out that this is typically a hint that those 2 positions are close in 3D. Part of the breakthrough has been our ability to decode that very effectively.

But what it implies is that, in the absence of that co-evolutionary landscape, the models don’t perform as well. When that information is available, maybe one could say the problem is somewhat solved from the perspective of structure prediction. When it isn’t, it’s much more challenging.

It’s also worth differentiating structure prediction and folding, because sometimes we conflate them a little bit. Folding is the more complex process of actually understanding how it goes from this disordered state into a structured-like state, and I don’t think we’ve made that much progress on that.

But the idea of going straight to the answer is something we’ve become pretty good at.

So there’s this protein that’s just a long chain, and it folds up.

Yeah.

And so we’re good at getting from that long chain, in whatever form it was originally, to the thing, but we don’t know how it necessarily gets to that state.

Yeah.

And there might be intermediate states that it’s in sometimes that we’re not aware of.

That’s right, and that also relates to our general ability to model the different…

Proteins are not static. They move. They take different—

Yeah.

…shapes based on their energy states, and I think we're also not that good at understanding the different states that the protein can be in, and at what frequency or probability.

Yeah.

So I think the 2 problems are quite related in some ways. There's still a lot to solve, but I think it was very surprising at the time that, even with these evolutionary hints, we were able to make such dramatic progress.

So I want to ask why the intermediate states matter. But first, I kind of want to understand why we care what proteins are shaped like.

4. Why Protein Shapes Matter

Yeah. Proteins are kind of the machines of our body. The way that all the processes that we have in our cells work is typically through proteins, sometimes through other molecules or intermediate interactions, and through those interactions we have all sorts of cell functions.

When we try to understand a lot of biology—how our body works, how diseases work—we often try to boil it down to, okay, what is going right in the case of our normal biological function, and what is going wrong in the case of the disease state? We boil it down to proteins and other molecules and their interactions.

So when we try predicting the structure of proteins, it's critical to have an understanding of those interactions. It's a bit like seeing the difference between having a list of parts that you would put in a car and seeing the car in its final form. Seeing the car really helps you understand what it does.

Right.

On the other hand, going to your question of why we care about how the protein folds, or how the car is made, to some extent, sometimes something goes wrong: there are cases of proteins misfolding in some diseases and so on. If we don't understand this folding process, we don't really know how to intervene.

There's this nice line in the AlphaFold 2 manuscript where they discuss why we're even hopeful that we can target the problem in the first place. There's this notion that, for proteins that fold, the folding process is almost instantaneous.

Mm-hmm.

Which is a strong signal that we might be able to predict this very constrained thing that the protein does so quickly. Of course, that's not the case for all proteins, and there's a lot of really interesting mechanisms in the cells. But I remember reading that and thinking, “Yeah, that's somewhat of an insightful point.”

I think one of the interesting things about the protein folding problem is that it used to be studied as a classical example of an NP-hard problem, and that's part of the reason why people thought it was impossible. There are so many different types of shapes that these amino acids could take, and so this grows combinatorially with the size of the sequence.

There used to be a lot of theoretical computer science thinking about and studying the protein-folding problem as an NP-hard problem. So it was very surprising also from that perspective, seeing machine learning make it so clear that there is some signal in those sequences through evolution, but also through other things that we as humans are probably not really able to understand, but that these models have learned.

And Andrew White—we were talking to him a few weeks ago—and he said that he was following the development of this and that there were actually ASICs that were developed just to solve this problem. Again, there were many millions of computational hours spent trying to solve this problem before AlphaFold.

Just to be clear, one thing that you mentioned was that there's this co-evolution of mutations that you see again and again in different species. So explain why that gives us a good hint that they're close to each other.

Yeah. Think of it this way: if I have some amino acid that mutates, it's going to impact everything around it in 3 dimensions, right?

Yeah.

So it's almost like the protein, through several probably random mutations in evolution, ends up figuring out that this other amino acid needs to change as well for the structure to be conserved.

Got it.

This whole principle is that the structure is probably largely conserved because there's this function associated with it. It's really different positions compensating for each other.

I see. Those hints in aggregate give us a lot of information about what is close to each other. Then you can start to look at what kinds of folds are possible given the structure, what the end state is, and therefore make a lot of inferences about what the actual total shape is.

Yeah, that's right. It's almost like you have this big, 3-dimensional valley where you're trying to find these low-energy states.

Yeah.

There's so much to search through that it's almost overwhelming. But these hints maybe put you in an area of the space that's already kind of close to the solution, maybe not quite there yet.

There's always this question of how much physics these models are learning versus pure statistics. I think one of the things I believe is that once you're in that approximate area of the solution space, the models have some understanding of how to get you to the lower-energy state.

So maybe you have some light understanding of physics, but maybe not quite enough to know how to navigate the whole space well.

Right. Okay.

So we need to give it these hints to get there.

So you get it into the right valley, and then it finds the minimum or something.

Yeah. Yeah.

Yeah.

One interesting explanation about how AlphaFold 3 works—which, of course, doesn't cover the entirety of what AlphaFold does—is one that I'm going to borrow from Sergey Chilingaryan at MIT. He sees AlphaFold, and the interesting thing about AlphaFold is that it's got this very peculiar architecture that we have since used. This architecture operates on these pairwise contacts between amino acids.

The idea is that probably the MSA gives you this first hint about what potential amino acids are close to each other.

MSA is mult—

Multiple sequence alignment. Exactly.

That's what he's talking about.

Exactly.

Yeah, yeah.

This evolutionary information.

Yeah.

From this evolutionary information about potential contacts, it's almost as if the model is running some kind of Dijkstra's algorithm, where it's decoding: okay, these have to be close. Then, if these are close and this is connected to this, this has to be somewhat close.

So you decode this, and that becomes basically a pairwise distance matrix. From this rough pairwise distance matrix, you decode the actual potential structure.

Interesting. So there's kind of 2 different things going on in the coarse-grained and then the fine-grained optimizations.

Exactly.

Interesting. Yeah. Very cool.

5. AlphaFold 3 Goes Generative

Yeah. You mentioned AlphaFold 3, so maybe this is a good time to move on to that. AlphaFold 2 came out, and it was, I think, fairly groundbreaking for this field. Everyone got very excited. A few years later, AlphaFold 3 came out.

For some more history, what were the advancements in AlphaFold 3? Then I think maybe after that we'll talk a bit about how it connects to Boltz.

Yeah. After AlphaFold 2 came out, Jeremy and I got into the field, along with many others. The clear problem that was obvious after that was, okay, now we can do individual chains. Can we do interactions? Interactions between different proteins, proteins with small molecules, proteins with other molecules—

So, quick: why are interactions important?

Interactions are important because, to some extent, that's the way these machines—that is, these proteins—have a function. The function comes from the way that they interact with other proteins and other molecules.

In the first place, the individual machines are often, as Jeremy was mentioning, not made of a single chain, but of multiple chains. These multiple chains interact with other molecules to give those their function.

On the other hand, when we try to intervene in these interactions, think about a disease, a biosensor, or many other ways we are trying to design molecules or proteins that interact in a particular way with what we would call a target protein or target—this problem after AlphaFold 2 became clear as one of the biggest problems in the field to solve. Many groups, including ours and others, started making contributions to this problem of trying to model these interactions. AlphaFold 3 was a significant advancement in modeling interactions. One interesting thing they were able to do was that, while some of the rest of the field had tried to model different interactions separately—how proteins interact with small molecules, how proteins interact with other proteins, and how RNA or DNA have their structures—they put everything together. They trained a very large model with many advances, including changes to some key architectural choices, and managed to get a single model that achieved this new state-of-the-art performance across all of these different modalities: protein–small-molecule interactions, which are critical to developing new drugs; protein–protein interactions; and understanding how proteins interact with RNA and DNA, and so on.

Just to satisfy the AI engineers in the audience, what were some of the key architectural and data changes that made that possible?

Yeah. One critical change that was not necessarily unique to AlphaFold 3, but that a few other teams, including ours, proposed, was moving from modeling structure prediction as a regression problem—where there is a single answer and you're trying to shoot for that answer—to a generative modeling problem, where you have a posterior distribution of possible structures and you're trying to sample from this distribution.

This achieves 2 things. One is that it starts to allow us to model more dynamic systems. As we said, some of these systems can actually take multiple structures, and so you can now model that by modeling the entire distribution. But on the other hand, from more core modeling questions, when you move from a regression problem to a generative modeling problem, you are really tackling the way that you think about uncertainty in the model in a different way.

So if you think about, “I'm undecided between different answers,” what happens in a regression model is that I'm going to try to make an average of those different answers I had in mind. When you have a generative model, what you're going to do is sample all these different answers and then maybe use separate models to analyze those different answers and pick out the best. So that was one of the critical improvements.

The other improvement is that they significantly simplified, to some extent, the architecture, especially the final model that takes those pairwise representations and turns them into an actual structure. That now looks a lot more like a traditional transformer than a very specialized equivariant architecture, which was what AlphaFold 3 had.

So this is a bitter lesson a little bit?

There is some aspect of a bitter lesson, but the interesting thing is that it's very far from being a simple transformer. This field is arguably one of the very few fields in applied machine learning where we still have architectures that are very specialized. Many people have tried to replace these architectures with simple transformers, and there is a lot of debate in the field, but I think most of the consensus is that the performance we get from the specialized architecture is vastly superior to what we get through a single transformer.

Another interesting thing on the modeling and machine-learning side, which I think is somewhat counterintuitive when you look at some of the other fields and applications, is that scaling hasn't really worked the same way in this field. Models like AlphaFold 2 and AlphaFold 3 are still very large models, but at the same time, in terms of parameters, they're actually not very big. They are definitely below 1 billion parameters.

If you hear these days in the LLM space, a model with less than 1 billion parameters, you'd think it can't do anything. But on the other hand, when you look at the computational cost of running these models, they're actually a lot more expensive than it is to run language models because, as Jeremy was saying, we go from having sort of quadratic operations to cubic operations. And so it's interesting how, right now in the field—and this may be related to having less data or needing more inductive biases—we have this ratio of the amount of computation to parameters that is much, much higher than in other places.

If I recall, AlphaFold 2 was, what, 70 million parameters? Something like that?

Yeah, it's something like that. It's quite small—less, around 100 or so. Yeah.

These decisions around triangular layers and this interesting equivariant architecture in AlphaFold 2 were really priors that baked in a lot of the physics of the system. And coevolution data is—I think people have argued that it is kind of almost like a database lookup of some sort.

Yeah, yeah.

It also sort of provides, in some sense, more parameters as well.

Yeah. I mean, it's definitely more about the amount of pure compute FLOPs, right, that is very high, and it's almost more like reasoning-based, maybe, than just information extraction.

It isn't just because of their reasoning capability; it's also because of the sheer quantity of information that they store. And I think here there's a little bit less of that, and I think it's more about decoding this input rather than maybe memorizing as much of it.

So is there a loop in the architecture that allows it to compute more per parameter? How does that work?

Part of it is exclusively this fact: instead of having operations that operate on a single chain, they operate on the pairwise representations. And so instead of having a quadratic number of interactions, you have a cubic number of interactions. That on its own leads you to have smaller representation sizes, but more representations, which leads to more FLOPs but fewer parameters.

Yeah.

On the other hand, there is also this idea, somewhat similar to reasoning, where you recycle this operation. Starting with AlphaFold 2, and also in AlphaFold 3, they have this interesting framework where you start—as we were discussing—the input to the model is sort of this initial understanding of the interactions, either from the evolution of the multiple sequence alignments, but also potentially from what we call templates, which are basically database lookups of similar structures.

And so how the model works is that it decodes these and tries to understand a good potential rough structure for the pairwise interaction. Then what you can do is basically do this recycling, where you feed this understanding back to the input of the model and then try to decode it again. People do this 3 or 4 times, and in some cases I've even tried to do it tens of times. And so you can see it as a very, very early version of reasoning.

So, AlphaFold 2, really cool. AlphaFold 3, really cool. But AlphaFold 3 came with a catch, and I think this catch was important for the development of Boltz and so on.

Yeah.

6. Boltz Answers The Open Source Gap

The catch was that it was an amazing Nature paper, but unfortunately they decided not to release the model. AlphaFold 2 was open source, and since then was used, I think the reported number is more than 1 million scientists.

AlphaFold 3, for commercial reasons—DeepMind has since spun off Isomorphic Labs, which is now trying to become a new pharmaceutical company—had decided to keep this model internal and only use it internally. Those of us in the field who were building on top of models like AlphaFold no longer had the base starting point to build on top of.

But even more importantly, everyone in both academic research and industry no longer had access to these incredible models, which were really useful for trying to understand biology but also for trying to develop new therapeutics. I decided to take the matter into our own hands and try to obtain a model that was of similar accuracy.

And so, largely using a lot of the information in the AlphaFold 3 manuscript, we went ahead and built Boltz-1, which was the first fully open-source model to approach the level of accuracy of AlphaFold 3. Along the way—and we can talk about it more—we realized that it was probably too ambitious to have this as an academic project. There were a lot of things that were missing, and so we decided to also start a public-benefit company to push this mission of democratizing access to these models, which we started with Boltz-1.

Quick interjection: I remember this. It was actually shocking how fast you got Boltz-1 out. It was just 2 or 3 months, right?

I think we started in late May, and it came in November—

Mm-hmm. Yeah.

—if I remember correctly. So, slightly longer, but yeah. It was relatively quick. For what it's worth, we were working on some similar ideas at the time. For example, this idea of having a diffusion model on top of a pairwise trunk was something that we were exploring independently.

When the paper came out, it was really clear, especially on the data pipelines, that there was so much we weren't really doing, and there was a lot to catch up on. But we were already in a place where we had some experience working with the data and with these types of models, and I think that put us in a good place to produce it quickly.

I would even say that we could have done it more quickly. The problem was that, for a while, we didn't really have the compute, so we couldn't train the model. We only trained the big model once—that's how much compute we had. We could only train it once. While the model was training, we were finding bugs left and right, a lot of them that I wrote.

Yeah.

I remember doing surgery in the middle of the run: stopping it, making the fix, relaunching it. We never actually went back to the start; we just kept training it with the bug fixes along the way.

Right.

Which was—

So it's impossible to reproduce now.

Yeah. That model has gone through such a curriculum that it's learned some weird stuff. But somehow, by some miracle, it worked out.

The other funny thing is that most of that model was trained through a cluster from the Department of Energy.

Yeah.

It was a shared cluster that many groups used, so we were basically training the model for 2 days, and then it would go back into the queue and stay there for a week.

Oh, man.

Oh.

It was pretty painful. Toward the end, with Devan, the CEO of Genesis, I was telling him a bit about the project and about our frustration with the compute. Luckily, he offered to help, and we got help from Genesis to finish the model. Otherwise, it probably would have taken a couple of extra weeks—

Wow.

—which—

Yeah.

Yeah.

Boltz-1.

How did that compare to AlphaFold 3? And then was there some progression from there?

I would say that both Boltz-1 and this other set of models that came around the same time were a big leap from the previous open-source models and were really approaching the level of AlphaFold 3. I would still say that, even to this day, there are some specific instances where AlphaFold 3 works better. One common example is antibody–antigen prediction, where AlphaFold 3 still seems to have an edge in many situations.

Obviously, these are somewhat different models. You run them and obtain different results, so it's not always the case that one model is better than the other. But in aggregate, especially at the time, AlphaFold 3 was still having a bit of an edge.

We should talk about this more when we talk about BoltzGen, but how do you know one model is better than the other? I make a prediction, you make a prediction—how do you know?

The great thing about structure prediction is that, a bit like what CASP was doing, there is a relatively straightforward way to evaluate the models. Once we go into the design space of designing new small molecules or new proteins, this becomes a lot more complex. You train a model on structures that had been released across the field up until a certain time. One thing we haven't talked about that was really critical in all this development is the PDB, the Protein Data Bank. It's this common resource, a basic shared database where every biologist can publish their structures.

We can train on all the structures that were put in the PDB until a certain date. Then we look at recent structures and ask which ones look pretty different from anything that was published before, because we really want to understand generalization. On these new structures, we evaluate all the different models.

And so you know when AlphaFold 3 was trained, and you intentionally train to the same date or something like that?

Exactly.

Right. Yeah.

That's the way you can somewhat easily compare these models. Obviously, that assumes you know the training set.

You've always been very passionate about validation. I remember DiffDock, and then there was DiffDock-L and DockGen. You've thought very carefully about this in the past. I think DockGen is a really funny story, and I don't know if you want to talk about that. It's interesting—

One of the amazing things about putting things open source is that we get a ton of feedback from the field. Sometimes we get great feedback from people who really like the model, but honestly, most of the time, people share where it doesn't work. To be honest, that may also be the most useful feedback.

Yeah.

At the end of the day, it's critical—and this is also true across other fields of machine learning—to set clear benchmarks. As you make progress on certain benchmarks, you need to improve the benchmarks and make them harder and harder. That's the progression of how the field operates.

The example of DockGen was that we published this initial model, called DiffDock, in my first year of my PhD. It was one of the early models to try to predict biomolecular interactions between proteins and small molecules, about a year after AlphaFold 2 was published.

On the one hand, on the benchmarks we were using at the time, DiffDock was doing really well and outperforming some of the traditional physics-based methods. On the other hand, when we started giving these tools to many biologists, one example being the group of Nick Polizzi at Harvard, we started noticing a clear pattern: for proteins that were very different from the ones the models were trained on, the models were struggling.

It seemed clear that this was probably where we should put our focus. So, first, with Nick and his group, we developed a new benchmark. Then we asked, “What can we change about the current architecture to improve this pattern of generalization?” That's what we're still doing today: asking where the model doesn't work, and then, once we have that benchmark, throwing every idea we have at the problem.

There's a lot of healthy skepticism in the field, which I think is great, and it's very clear that there are a ton of things the models don't work well on. But one thing that's probably undeniable is the pace of progress and how much better we're getting every year.

If you assume any constant rate of progress moving forward, I think things are going to look pretty cool at some point in the future.

ChatGPT was only 3 years ago.

Yeah, it's wild, right?

What?

Yeah, it's one of those things—even being in the field, you don't see it coming. Hopefully, we'll continue to have as much progress as we've had over the past few years.

So this is maybe an aside, but I'm really curious. You get this great feedback from the community by being open source, right? My question is partly: if you open-source, then everyone can copy what you did, but it's also maybe about balancing priorities, right?

All my customers are saying, “I want this.” There are all these problems with the model, but my customers don't care, right? How do you think about that?

Yeah.

So I would say a couple of things. One is that part of our goal with Boltz—and this is also established as the mission of the public benefit corporation that we started—is to democratize access to these tools. But one of the reasons why we realized that Boltz needed to be a company, and couldn’t just be an academic project, is that putting a model on GitHub is definitely not enough to get chemists and biologists across academia, biotech, and pharma to use your model in their therapeutic programs. A lot of what we think about at Boltz, beyond just the models, is all the layers that come on top of the models to get from those models to something that can really enable scientists in the industry.

And so that goes into building the right kind of workflows that take in, for example, the data and try to answer directly the problems that the chemists and biologists are asking, and then also building the infrastructure. This is to say that even with models fully open, we see a ton of potential for products in this space. The critical part about a product is that even with an open-source model, running the model is not free. As we were saying, these are pretty expensive models, and especially—and maybe we’ll get into this—these days we’re seeing pretty dramatic inference-time scaling of these models—

Yes.

—where the more you run them, the better the results are. But you start getting to a point where compute and compute cost become critical factors. Putting a lot of work into building the right kind of infrastructure, building the optimizations, and so on, really allows us to potentially provide a much better service for open-source models.

That is to say, even though we’re building the product and can provide a much better service, I do still think—and we will continue to put a lot of our models open source—that the critical role of open-source models is helping the community progress on the research, from which we all benefit. We’ll continue, on the one hand, to put some of our base models open source so that the field can build on top of them. As we discussed earlier, we learn a ton from the way that the field uses and builds on top of our models.

But then we’ll try to build a product that gives the best experience possible to scientists, so that a chemist or biologist doesn’t need to spin up a GPU and set up our open-source model in a particular way, but can just use it. Even though I am a computer scientist and machine-learning scientist, I don’t necessarily take an open-source LLM and try to spin it up; I might just open the ChatGPT app or Claude Code and use it as an amazing product. We want to give the same experience to scientists around the world.

I heard a good analogy yesterday: a surgeon doesn’t want the hospital to design a scalpel, right? The surgeon just buys the scalpel.

You wouldn’t believe the number of people who, even in my short time between AlphaFold 3 coming out and the end of the PhD, would reach out just for us to run AlphaFold 3 for them. Or things like that, just because Boltz, in our case—because it’s not that easy to do that computation if you’re not a computational person. I think part of the goal here is also that we continue to build an interface for computational folks, but that the models are also accessible to a larger, broader audience. That comes from good interfaces and things like that.

I think one really interesting thing about Boltz is that with its release, you didn’t just release a model, but you created a community. Did that community—it grew very quickly. Did that surprise you? And what is the evolution of that community, and how has that fed into Boltz—

If you look—

—as a company?

If you look at its growth, it’s—

Yeah.

—it’s very much like when we release a new model: there’s a big jump. But, yeah, it’s been great. We have a Slack community that has thousands of people on it, and it’s actually self-sustaining now, which is really nice. It’s almost overwhelming to answer everyone’s questions and help with the few people we had, but it ended up that people would answer each other’s questions and help one another. The Slack has been self-sustaining, and that’s been really cool to see.

For the Slack part, but then also obviously on GitHub as well, we’ve had a nice community. We also aspire to be even more active on it than we’ve been in the past 6 months, which has been a bit challenging for us. But the community has been really great. There are also a lot of papers that have come out with new evolutions on top of Boltz, and that surprised us to some degree because there are a lot of models out there. People converging on Boltz was really cool.

I think it also speaks to the importance of trying to put a lot of emphasis, when you put code out, on making it as easy to use as possible—something we thought a lot about when we released the codebase. It’s far from perfect, but—

Do you think that was one of the factors that caused your community to grow—just the focus on making it easy to use and accessible?

I think so, yeah. We’ve heard it from a few people over the years now, and some people still think it should be a lot nicer. They’re right. But, yeah, I think it was, at the time, maybe a little bit easier than other things.

The other thing that I think led to the community, and to some extent to the trust in the community and in what we put out, is the fact that it’s not really been just one model. Maybe we’ll talk about it: after Boltz-1, there were another couple of models released or open-sourced soon after. We continued the open-source journey and released Boltz-2, where we were not only improving structure prediction, but also starting to do affinity predictions—understanding the strength of the interactions between these different molecules, which is a critical component, a critical property that you often want to optimize in discovery programs.

More recently, we also released a protein-design model. We’ve been building this suite of models that come together and interact with one another. There is almost an expectation that we take very much to heart always having, across the entire suite of different tasks, the best or among the best models out there, so that our open-source tool can be the go-to model for everybody in the industry.

I really want to talk about BoltzGen. But before that, one last question in this direction: was there anything about the community that surprised you? Were there any moments when someone was doing something and you thought, “Why would you do that? That’s crazy,” or, “That’s actually genius, and I never would have thought about that”?

We’ve had many contributions. Some of the interesting ones—I mean, we had this one individual who wrote a complex GPU kernel for part of the architecture. The funny thing is that that piece of the architecture had been there since AlphaFold 2.

I don’t know why it took Boltz for this person to decide to do it, but that was a really great contribution. We’ve had a bunch of others, like people figuring out ways to hack the model to do cyclic peptides. I don’t know if any other interesting ones come to mind.

One cool one—and this was initially proposed as a message in the Slack channel by Tim O’Donnell—was basically that there are some cases, especially antibody–antigen interactions, where the models don’t necessarily get the right answer. What he noticed is that the models were somewhat stuck predicting the antibody to interact with the part of the antigen that was incorrect.

And so he basically ran the experiments. In this model, you can condition it; basically, you can give it hints. He gave random hints to the model: “Okay, you should bind to this residue. You should bind to the 1st residue, or you should bind to the 11th residue, or you should bind to the 21st residue,” or basically every 10 residues, scanning the entire antigen.

Residues are the—

The amino acids.

The amino acid, yeah.

So the 1st amino acid, the 11th amino acid, and so on. It’s sort of like doing a scan, conditioning the model to predict all of them, looking at the confidence of the model in each of those cases, and taking the top. It’s a somewhat crude way of doing inference-time search, but surprisingly, for antibody-antigen prediction, it actually helped quite a bit.

And so there are some interesting ideas that, as you’re developing the model, you say, “Wow, why would the model be so dumb?” But it’s very interesting, and that leads you to start thinking, “Okay, can I do this—not with this brute force, but in a smarter way?” We’ve also done a lot of work in that direction.

And that speaks to the power of scoring. We’re seeing that a lot. I’m sure we’ll talk about it more when we talk about BoltzGen, but our ability to take a structure and determine that that structure is good—somewhat accurate—whether that’s a single chain or an interaction, is a really powerful way of improving the models.

If you can sample a ton and assume that, if you sample enough, you’re likely to have the good structure, then it really just becomes a ranking problem. Part of the inference-time scaling that Gabriele was talking about is very much that. The more we sample, the more the ranking model ends up finding something it really likes. And so I think our ability to get better at ranking is also what’s going to enable the next big breakthroughs.

Interesting.

I guess there’s—my understanding is that there’s a diffusion model, and you generate some things, and then—

Yeah.

Then you rank them using a score, and then— I guess it’s just what you said, right? Can you talk about those different parts?

7. BoltzGen Designs New Proteins

Yeah. So first of all, one of the critical beliefs that we had, also when we started working on Boltz-1, was that structure-prediction models are somewhat our field’s version of foundation models. They learn about how proteins and other molecules interact, and then we can leverage that learning to do all sorts of other things. With Boltz-2, we leveraged that learning to do affinity predictions—to understand, if I give you this protein and these small molecules, how tightly is their interaction?

For BoltzGen, what we did was take the foundation models and fine-tune them to predict entirely new proteins. The way that works is that, instead of filling in an actual sequence for the protein you’re designing, you fill in a set of blank tokens, and you train the models to predict both the structure of that protein and, with the structure, what the different amino acids of that protein are.

Basically, the way BoltzGen operates is that you feed it a target—a protein that you may want to bind to, or DNA or RNA—and then you feed it the high-level design specification of what you want your new protein to be. For example, it could be an antibody with a particular framework, a peptide, or many other things.

And that’s with natural language or with this—

And that’s basically prompting, and we have this sort of spec that you specify.

Okay.

You feed this spec to the model, and then the model translates it into a set of tokens, a set of conditioning for the model, a set of blank tokens. Then it decodes, as part of the diffusion model, a new structure and a new sequence for your protein. Then we take that and, as Jeremy was saying, try to score it and see how good of a binder it is to that original target.

You’re basically using Boltz-2 to predict the folding and the affinity to that molecule, and that gives you a score. Is that right?

Exactly.

Yeah.

So you use this model to predict the structure, and then you do 2 things. One is that you predict the structure with something like Boltz-2, and then you compare that structure with what the model predicted—what BoltzGen predicted.

This is sort of called consistency in the field. You want to make sure that the structure you’re predicting is actually what you’re trying to design, and that gives you much better confidence that it’s a good design. That’s the first filtering.

The second filtering that we did as part of the BoltzGen pipeline, and that was released, is that we look at the confidence that the model has in the structure. Now, unfortunately, going to your question about predicting affinity, confidence is not a very good predictor of affinity.

One of the things that we’ve made a ton of progress on since we released BoltzGen—and we have some new results that we’re going to announce soon—is the ability to get much better hit rates. Instead of trying to rely on the confidence of the model, we’re actually trying to predict the affinity of that interaction directly.

Okay. Just backing up a minute: your diffusion model actually predicts not only the protein sequence but also its folding?

Exactly.

And actually, one of the big differences that we made compared to other models in the space—and there were some papers that had already done this before, but we really scaled it up—was somewhat merging structure prediction and sequence prediction into almost the same task.

The way BoltzGen works is that the only thing you’re doing is predicting the structure. The only supervision is supervision on the structure. But because the structure is atomic, and the different amino acids have different atomic compositions, from the way that you place the atoms, we also understand not only the structure that you wanted, but also the identity of the amino acid that the model believed was there.

And so, instead of having these 2 supervision signals—1 discrete, 1 continuous—that don’t interact particularly well, we build an encoding of sequences and structures that allows us to use exactly the same supervision signal that we were using for Boltz-2, which is largely similar to what AlphaFold 3 proposed. It’s very scalable, and we can use it to design new proteins.

Oh, interesting.

Maybe a quick shout-out to Hannes Stärk on our team, who did all this work. Yeah.

Yeah, that was a really cool idea. Looking at the paper, there’s this encoding where you just add a bunch of atoms that can be anything, and then they get rearranged and basically plopped on top of each other, so that—and then that encodes what the amino acid is. There’s a unique way of doing this. That was such a cool, fun idea.

I think that idea had existed before. I think it wasn’t—

Yeah, there were a couple of papers that had proposed this, and—

Yeah, yeah.

—and Hannes really took it to a large scale.

Yeah.

In the paper, a lot of it is dedicated to actually validating the model. In my opinion, all the people we talk about basically feel that this sort of wet-lab—or whatever the appropriate real-world validation is—is the whole problem, almost. Or not the whole problem, but a big, giant part of the problem.

Can you talk a little bit about the highlights from there? To me, the results are impressive both from the perspective of the model and also just the effort that went into the validation by a large team.

First of all, I think I should start by saying that both when we were at MIT, where Tommi Jaakkola was working in Barzilay’s lab, as well as at Boltz, we are not a biolab, and we are not a therapeutic company.

And so, to some extent, we were first forced to look outside our group, our team, to do the experimental validation. One of the things that Hannes and the team pioneered was the idea: can we go not only to a specific group and try to find a specific system, maybe overfit a bit to that system, and try to validate, but how can we test these models across a very wide variety of different settings? Protein design is such a wide task, with all sorts of different applications, from therapeutics to biosensors and many others. Can we get a validation that goes across many different tasks?

And so he basically put together, I think, something like 25 different academic and industry labs that committed to testing some of the designs from the model. Some of this testing is still ongoing, and they're giving results back to us in exchange for hopefully getting some new, great sequences for their task. He was able to coordinate this very wide set of scientists, and already in the paper, I think we shared results from 8 to 10 different labs, showing results from designing peptides to target ordered proteins, peptides targeting disordered proteins, and proteins that bind to small molecules. We showed results of designing nanobodies across a wide variety of different targets. And so that gave the paper and the model a lot of validation that was wide.

And so, would those be therapeutics for those animals, or are they relevant to humans as well?

They're relevant to humans as well. Obviously, you need to do some work into, quote-unquote, humanizing them, making sure that they have the right characteristics so they're not toxic to humans and so on. There are some approved medicines on the market that are nanobodies.

There's a general pattern, I think, in trying to design things that are smaller. It's easier to manufacture. At the same time, that comes with potentially other challenges, like maybe a little bit less selectivity than if you have something that has more heft.

Yeah.

There's a big desire to design mini-proteins, nanobodies, and small peptides that are just great drug modalities.

Okay. I think we left off talking about validation in the lab, and I was very excited about seeing all the diverse validations that you've done. Can you—

Yeah.

—go into some more detail about them?

Yeah.

The specific ones.

Yeah. The nanobody one, I think we did—what was it? 15 targets? Is that correct?

14.

14 targets. Typically, the way this works is that we make a lot of designs, on the order of tens of thousands, and then we rank them and pick the top N. In this case, N was 15 for each target, and then we measure the success rates, both in terms of how many targets we were able to get a binder for and, more generally, out of all of the binders that we designed, how many actually proved to be good binders.

Some of the other ones involved a cool example where there was a small molecule that a designer protein binds to. That has a lot of interesting applications. For example, as Gabriele mentioned, biosensing and things like that.

Okay.

We had a disordered protein, I think you mentioned also. Those were maybe some of the highlights.

Yeah. I would say that the way we structured some of those validations was, on the one hand, that we had validations across a whole set of different problems that the biologists we were working with came to us with. We were trying, for example, in some of the experiments, to design peptides that would target RexC, which is a target involved in metabolism. We had a number of other applications where we were trying to design peptides or other modalities against some other therapeutically relevant targets. We designed some proteins to bind small molecules.

Some of the other testing that we did was really trying to get a broader sense of how the model works, especially when tested on generalization. One of the things that we found in the field was that a lot of the validation, especially outside of the validation that was done on specific problems, was done on targets that have a lot of known interactions in the training data. It's always a bit hard to understand how much these models are really just regurgitating or trying to imitate what they've seen in the training data, versus really being able to design new proteins.

One of the experiments that we did was to take 9 targets from the PDB, filtering to things where there was no known interaction in the PDB. Basically, the model had never seen this particular protein bound, or a similar protein bound to another protein. There was no way that the model, from its training set, could say, “Okay, I'm just going to—”

Like, tweak something.

“—tweak something and just imitate this particular kind of interaction.”

We took these 9 proteins, worked with Adaptyv, a CRO, and basically tested 15 mini-proteins and 15 nanobodies against each one of them. The very cool thing that we saw was that, on two-thirds of those targets, we were able to get nanomolar binders from these 15 designs. Nanomolar, roughly speaking, is just a measure of how strongly the interaction is.

And roughly speaking, a nanomolar binder is approximately the binding strength that you need for a therapeutic.

Okay.

8. Boltz Lab Turns Models Into Workflows

So, maybe switching directions a bit, Boltz Lab was just announced this week. Or was it last week?

Yeah.

This is your first product, I guess, if you want to call it that. Can you talk about what Boltz Lab is and what you hope people take away from this?

Yeah. As we mentioned at the very beginning, the goal with the product has been to address what the models don't do on their own, and there are largely 2 categories there. I'll split it into 3.

The first one: it's one thing to predict a single interaction, for example, a single structure. It's another to very effectively search a design space to produce something of value. What we found building this product is that there are a lot of steps involved in that, which we need to accompany the user through.

One of those steps, for example, is the creation of the target itself. How do we make sure the model has a good enough understanding of the target so we can design something? There are all sorts of tricks that you can do to improve a particular structure prediction. That's the first stage.

Then there's the stage of designing and searching the space efficiently. For something like BoltzGen, for example, you design many things and then rank them. But for a small molecule, the process is a little more complicated. We actually need to also make sure that the molecules are synthesizable.

The way we do that is that we have a generative model that learns to use appropriate building blocks, such that it can design within a space that we know is synthesizable. There's this whole pipeline of different models involved in being able to design a molecule. That's been the first thing.

We call them agents. We have a protein agent, and then we have a small-molecule design agent. That's really at the core of what powers the Boltz Lab platform.

So, are these agents a language-model wrapper, or are they just your models that you're calling agents because they perform a function?

No. They're more of a recipe, if you wish. I think we use that term because of the complex pipelining and automation that goes into all this plumbing. That's the first part of the product.

The second part is the infrastructure. We need to be able to do this at very large scale for any one group that's doing a design campaign.

Let's say you're designing 100,000 possible candidates to find the good one, right? That's a very large amount of compute. For a small molecule, that's on the order of a few seconds per design. For proteins, it can be a bit longer. Ideally, you want to do that in parallel; otherwise, it's going to take you weeks. We've put a lot of effort into our ability to have a GPU fleet that allows any one user to do this kind of large parallel search.

So you're amortizing the cost over your users, basically.

Exactly.

Yeah.

Exactly. To some degree, using 10,000 GPUs for a minute is the same cost as using 1 GPU for God knows how long, right? You might as well try to parallelize if you can. A lot of work has gone into making that very robust, so that we can have a lot of people on the platform doing that at the same time.

The third one is the interface, and the interface comes in 2 shapes. One is in the form of an API, and that's really suited for companies that want to integrate these pipelines, these agents, directly into existing workflows or existing user interfaces that they have. We're already partnering with a few distributors that are going to integrate our API.

Then the second part is the user interface. We've put a lot of thought into that as well. This is what I mentioned earlier, this idea of broadening the audience. That's what the user interface is about, and we've built a lot of interesting features into it, for example, for collaboration. When you have potentially multiple medicinal chemists going through the results and trying to pick out which molecules we're going to go and test in the lab, it's powerful for them to be able to, for example, each provide their own ranking and then do consensus building. There are a lot of features around launching these large jobs, but also around collaborating on analyzing the results.

Yeah.

And so Boltz Lab is a combination of these 3 objectives into 1 cohesive platform.

Who is this accessible to?

Everyone. You do need to request access today. We're still ramping up usage, but anyone can request access. If you're an academic, in particular, we provide a fair amount of free credit so you can play with the platform. If you're a startup or a biotech, you may also reach out, and we'll typically hop on a call just to understand what you're trying to do and provide a lot of free credits to get started. Of course, with larger companies, we can deploy this platform in a more secure environment. Those are more custom deals that we make with partners.

That's part of the ethos of Boltz. I think this idea of servicing everyone, and not necessarily going after just the really large enterprises, starts with open source. It's also a key design principle of the product itself.

Yeah.

One thing I was thinking about with regard to infrastructure: in the LLM space, the cost of a token has gone down by, I think, a factor of 1,000 or so over the last 3 years, right?

Yeah.

Is it possible that, essentially, you can exploit economies of scale and infrastructure so that you can make it cheaper to run these things yourself than for any person to roll their own system?

Oh, 100%. We're already there. Running Boltz on our platform, especially at large scale, is considerably cheaper than it would probably take anyone to put the open-source model out there and run it. On top of the infrastructure, one of the things that we've been working on is accelerating the models. Our small-molecule screening pipeline is 10x faster on Boltz Lab than it is in the open-source implementation, and that's also part of building a product—something that scales really well. We really wanted to get to a point where we could keep prices very low, in a way that it would be a no-brainer to use Boltz through our platform.

Good. How do you think about validation of your agentic systems?

Yeah.

Because, as you were saying earlier, AlphaFold-style models are really good at, let's say, monomeric proteins where you have co-evolution data. But now suddenly the whole point of this is to design something that doesn't have co-evolution data, something that is really novel. So now you're basically leaving the domain that you thought was—

Right.

—that you were good at. So how do you validate that?

Yeah, I agree completely, but there's obviously a ton of computational metrics that we rely on, but those only take you so far. You really have to go to the lab and test: with Method A and Method B, how much better are we? How much better is my hit rate? How much stronger are my binders? It's not just about hit rate; it's also about how good the binders are. There's really no way around that. I think we've really ramped up the amount of experimental validation that we do so that we can really track progress as scientifically sound as possible. I don't know if there's anything else.

I think one thing that's unique about us, and maybe companies like us, is that we're not working on just a couple of therapeutic pipelines with validation focused on them. When we do an experimental validation, we try to test it across tens of targets. That way, on the one hand, we can get a much more statistically significant result, and that really allows us to make progress from the methodological side without being steered by overfitting on any one particular system.

Of course, we choose—we always try to choose—targets and problems that are at the frontier of what's possible today. You don't want something too easy, and you don't want something too hard; otherwise, you're not going to see progress. This is a somewhat evolving set of targets. We talked earlier about the targets that we looked at with BoltzGen, and now we're trying even harder targets, both for small molecules and proteins. We try to keep ourselves on the boundary of what's possible.

Do you have infrastructure, or do you just have a lot of different partnerships with academic labs and keep pushing on these and driving them forward?

We do this partially through academic labs. More and more, we do this through CROs, just because, to some extent, we need replicability—often going after the same targets multiple times and seeing the progress from 1 month to the next.

Yeah.

And speed.

Speed of execution.

What happens if you start getting a bunch of really strong binders against therapeutic targets? What do you do?

Release them.

Yeah. Put them online.

Oh, you release them as open source?

Yeah. When we say we have no interest in making drugs, we're serious.

Yeah.

With the academic labs, basically, they keep them and do whatever they want with them. With the CROs, so far, we've been releasing them.

I will also say—and I think this has been a bit of the issue that I have with some of the things that have been said in the field—that when we say we design new proteins, or we say we design new molecules that go and bind these particular targets, we should be very clear: these are not drugs.

Yeah.

These are not things that are ready to be put—

Yes. Yeah.

—into a human. There is still a lot of development that goes with it. To us, we see ourselves as building tools for scientists. At the end of the day, it really relies on the scientist having a great therapeutic hypothesis and then pushing through all the stages of development. We try to build tools that can accompany them in that journey.

It’s not like a magic box where you can just turn it and get—

Get FDA-approved drugs.

FDA-approved drugs. Yeah.

But actually, that brings up an interesting question that I’ve been wondering about. Do you guys see yourself staying in this, for lack of a better way of saying it, layer? Or do you think that you’ll start to, in a physical sense, look at different layers of the virtual cell, so to speak?

There’s also the development process that goes sort of like design, preclinical, clinical, approval, and thinking about improving performance throughout that process based on the designs. Is that a direction that you guys are pushing?

Yeah. So one of the things, as Jeremy said, is that we are not a therapeutic company, and we want to stay that way—not be a therapeutic company—and always be at the service of all the different companies, including therapeutic companies, that we serve. To some extent, that does mean that we need to try to go deeper and deeper in getting these models better and better.

One of the things that we’re doing, along with many others in the field, is that now we’re starting to have models that are good both for small molecules and for proteins, and that can design binders—relatively tight binders. We’re starting to look at all these other properties. They’re called developability or ADME, which we care about when developing a drug, and we’re trying to figure out whether we can design for them from the get-go.

The thing about those properties is that, for some of them, you need to start having an understanding of the cell. That’s, on the one hand, why we need that understanding. But also, the way that we think about all the different and complex diseases is that these models and tools we’re building need to have a good understanding of biomolecular interactions.

At the same time, every disease is often unique, and every therapeutic hypothesis is unique. You may want something that needs to hit a particular target in a virus in a particular way, but you may not know exactly what way you want to do that. In the first set of designs, you’re going to try to target different epitopes in different ways. Then you’re going to test them in the lab, maybe directly in vivo, and see which ones work and which ones don’t.

You need to bring those results back into the models, and then the models can start to have a wider understanding—not just of the biophysics of the antibodies interacting with that target, but also of how that is shaped within the entire cell. First of all, that means, on the one hand, that we need these loops. This is also partially how we designed the platform to be.

But it also means that we need to start understanding more and more higher-level things. I wouldn’t say that we’re working in any way on a virtual cell like others are, but we’re definitely thinking very deeply about how the way that we target certain proteins interacts with pathways that are existing in the cell.

One question that has come up is that you talk a lot about user interface and so on, and I think this is really important. My experience dealing with medicinal chemists, when you give them machine-learning models, is that they’re the most superstitious, skeptical, pseudo-religious people I’ve ever talked to when it comes to doing science.

Sorry for the medicinal chemists listening.

Yeah. They’re amazing. I’ve worked with some spectacular medicinal chemists who just pull magic out of their hat again and again, and I have no idea how they do it. But when you bring them a machine-learning model, it can sometimes be quite tricky to get them to deal with it. How has your interaction been with this, and how have you thought about building Boltz Lab to work with the skeptics?

One of the great value unlocks for us and for our product has been bringing a medicinal chemist onto the team. His name is Jeffrey. On day 1, he obviously had a lot of opinions about many of the ways that we should change both the way the agents worked and the way the platform worked.

But it’s been really amazing, once we started shaping the platform in a better way with his feedback, how we went from a fair amount of skepticism from him to him actually using a lot more compute than any of our computational folks on the team. At times, he’s running all these sorts of hypotheses: “Okay, maybe I can hit this protein this particular way. I can hit it in that way. Actually, let me look at this particular molecular space. Let me try to optimize for these particular interactions.”

He ends up running several screens in parallel, using hundreds of GPUs on his own. It’s been pretty incredible to see how the way that I was thinking about a problem—“Okay, you’re just trying to design a binder, a small molecule to a particular protein”—is much simpler than the way that he thinks about it. He thinks about it much more deeply, trying all these different things and these different hypotheses.

Once he gets the results from the model, he doesn’t just take the top 15. He really looks over them and tries to understand the different things. Then, when we select maybe some designs to bring forth, he has something where both the models understand that something’s good, but he himself does as well. That’s why we also built the platform to be an interface for this kind of chemist and also a collaborative experience.

At the end of the day, I think that for people to be convinced, you have to show them something they didn’t think was possible. Until you have that aha moment, I think the skepticism will remain. But every once in a while, there’s a result that really surprises people, and then it’s like, “Oh, wow. Okay, this is actually—I can do something with this.”

So you just get it in their hands, have them try it out, and they’ll be convinced. Yeah.

Yeah, or maybe once the lab results come back.

Or maybe one of their colleagues is convinced.

Yeah.

I think it takes going to the lab—

Yeah.

I think at some point. There’s no avoiding that. As beautiful as the platform can be, and as nice as the molecules that the model predicted might look, I think what really convinces people is hits.

Yeah.

Yeah.

You see the results, and yeah.

Exactly. Yeah.

Cool. Thank you for taking the time to chat with us.

Yeah. It’s been really interesting.

Um, is there anything that you would like your audience to know?

I mean, first of all, we’re just getting started, continuing to build the team, and so definitely always looking for great folks both on the software side, machine-learning side, but also scientists to join the team and help us shape—

On the infrastructure side too? Like—

Indeed.

If you think that if you want a new challenge, because this is not just next-token prediction, this is really a new engineering challenge—

Exactly.

—that hasn’t been done before.

If you, no matter how much experience you have with biologists and chemistry, if you want to come help us shape what biology and chemistry hopefully will look like in five, ten years, we’d love to hear from you. Go to Boltz.bio and come join the team.

Cool. Thank you.

Awesome.

Thank you so much.

Thanks so much.

Thank you.