Biohub：生物学的未来是开源的——联合创始人 Mark Zuckerberg、Priscilla Chan 与科学负责人 Alex Rives

We just want to give tools to the whole scientific community.

We want to understand how biology works. I want to understand the genetics of this person. I want to understand the risks they have for different illnesses. My goal is to be able to treat the individual as an individual, understand the mechanisms, and be able to intervene.

We'll have a bigger impact by getting this in more scientists' hands quicker by doing it as open-source projects instead. It's not just like there's some factory somewhere that you can pay to produce the data. You actually need to invent new, novel scientific approaches.

The theory isn't that we're going to cure the diseases. We're not. It's that we want to help accelerate the pace of progress for the whole scientific field.

We folded over 1.1 billion proteins and predicted their structures, and we didn't design a model for antibodies. We didn't design a model to be able to bind one particular target. We just designed a model that could understand proteins.

If we could design a protein to actually change the physiology, then we can actually cure someone.

Today on No Priors, we're joined by Mark Zuckerberg, Priscilla Chan, and Alex Reeves. We'll be talking about Biohub and all their various efforts to apply AI at scale to develop world models of cells and different levels of interaction across biology.

Mark, Priscilla, thank you for doing this.

Yeah, thanks for having us.

Great to be here.

This is fun.

Alex, congratulations on new missions.

Thank you.

You guys made Biohub your primary philanthropic effort and then committed $500 million to this virtual biology initiative. Can you tell us a little bit about why you did that, and how you went from, “We should fund this,” to, “This is who we are”?

1. Biohub Builds Scientific Tools

We're super excited about Biohub in its current form. We feel like it's a really good fit for who we are, what we bring to the table, and what we can achieve together. But this work started 10 years ago when we were thinking about how we could give back.

Mark wanted to build an organization that could cure, prevent, and manage all disease by the end of the century. And we had a series of hilarious meetings with scientists where famous Nobel Prize-winning scientists were just laughing at us.

Was that your starting line? “We're just going to cure all disease”?

No, no. And to be clear, we don't think that we're going to be the ones curing the diseases. Our goal was always to build tools that could accelerate the whole scientific field, that way the scientific field collectively could cure all the diseases. But still—

But still, people laughed at us.

I mean, people thought that by the end of the century was a stretch. Now I think it's too conservative.

And so we kept saying, “Okay, well,” and had a series of funny, awkward educational conversations. We'd say, “Okay, but why? Why do you think it's impossible?” We were just being the people in the room who said, “Well, I don't know why. You tell me.”

Finally, we got people to say, “Fine, if you really must know.” And we'd say, “You know, we do. It seems important.” They said, “Well, we work in silos. When you publish, information doesn't get shared. It gets locked up for long periods of time, and we don't have tooling.”

They gave the example: “We build a great tool by 1 postdoc in a lab, and it lives on their computer, and when they graduate, the tool is gone.” What we heard was that it was very hard to build shared tools to move science faster and build a shared knowledge base to quickly move science faster.

That's where we began thinking about, if those are the problems, what can we contribute?

Yeah, the original Biohub model was basically to focus on long-term tool development by bringing together engineers and scientists across multiple universities to focus on long-term tool development. It basically worked.

We started off with CZI doing a number of different things, and over time, we just felt like the science piece was really working. We kept investing more and more in it until now it is basically the primary and main thing that we're doing.

We've expanded the original San Francisco Biohub to a handful at this point. There's New York. There's Chicago. The real focus and the unifying theme at this point is the virtual biology initiative, around taking the unique data sets that are able to be generated in order to model, effectively starting with the smallest pieces of proteins, but then eventually cells and whole biological systems.

That's kind of how we've evolved: this idea that some of this is an AI problem. You want to build a frontier AI lab, but you need to couple that with a frontier biology effort that can do the work of understanding and getting the data that you need to actually be able to build these models.

Unlike language models, where there's a lot of data out there on the internet, that's not really the case with biology. There are obviously a bunch of different data sets that exist that academia and scientists have generated over the decades, but a lot of the stuff that I think we want to put into this doesn't exist.

You want to be able to visualize things that people haven't been able to see before, which is why we're doing the imaging work. You want to be able to record things that are going on inside the body, which is why we're doing the cellular engineering work. Or you want to be able to measure things like inflammation in ways that haven't been possible, which is why Chicago Biohub is focused on building those kinds of devices.

That will fundamentally create new types of data sets that will allow new types of models. I think that's just a very exciting thing. Going back to what you were saying, if the scientific field primarily needs tool development, that is going to empower scientists across the field to be able to do their work faster. That's what we think we can provide through this kind of long-term focus on tool development.

But I think there's a fun through line from where we started to the work that Alex is driving now: our very first request for applications, RFA, was around single-cell sequencing. We wanted to look at the RNA that is transcribed in individual cells.

Mm-hmm.

That was possible, but it was still pretty early on in understanding how different cells were expressing their DNA. At the beginning, we were just funding methods, getting people to describe how to do it so that others could share that methodology.

Then that became us funding the Human Cell Atlas, which is now one of the largest databases of single-cell transcriptomes. It was getting hard for scientists to annotate the data, so we built Cell by Gene, which was a very simple annotation tool that scientists could use to make use of that data.

Then a community came around Cell by Gene, built around Cell by Gene, and started contributing more and more data that we had nothing to do with creating, funding, or making happen in the world. Now Cell by Gene is a corpus of knowledge that a lot of the transcriptomic-based models are based on, and it's used regularly by the scientific community.

But there are always critiques: “This is just stamp collecting. You're just gathering bits of knowledge—well, sorry, bits of data—and we're not going to be able to pull scientific knowledge, wisdom, and insights out of it.” We didn't have an answer for a while.

Then imagine our delight when large language models became a huge topic of conversation that could make sense of large amounts of data. For me, it was: What if we could actually understand how biology worked? Move it from a discovery-based science to an engineering-based science, where we could systematically understand how living beings and living cells worked, and be able to understand why things go wrong.

When we saw that moment, we were like, “This is it. Something really big could happen here.”

Alex, you started at Metafair, but you'd assembled a team at EvolutionaryScale, raised venture, and were making progress on your models. What was the pitch from Mark and Priscilla that made you say, “That's actually the right way to go after the mission”?

Well, for me, it was the moment when I understood that they really saw this as an integration of frontier AI and frontier biology. I think I had developed the conviction that this is really a new era of science that's just beginning in terms of what's going to be possible with artificial intelligence.

We're in the age of information theory at scale, and we have these systems that can basically predict the next token, and they can learn world models from that.

Mm-hmm.

They can learn biology from the data. And so I think it was really clear that, to build that next institution for the next era, you would really need to have frontier artificial intelligence.

You would have to have frontier biology. You would need to start to put those things in feedback and really have models that are learning from the biology. You'd need the right scale and the right people, and so this just really felt, I think, like the way to do that.

2. Biology Needs Hierarchical Models

There's a variety of different models that you all have been working on. I think it's interesting because some of the earliest breakthroughs in biology were things like AlphaFold, where it was a Google model that showed you could do protein folding at scale in a really interesting way that people didn't realize was very tractable. This was before the really big transformer waves that came later.

Then you're working on a variety of different things at different scales, right? You're doing individual molecular modeling and protein folding. You're doing cell-based stuff. You're thinking about interrogating larger-scale systems in biology. How well do you think that extends from the micro to the macro? You mentioned almost starting with building blocks and building up, but modeling cellular behavior is very different from modeling protein folding. The data is very different. The modeling is different. Do you think it's all similar in terms of just being data that you train on, or do you think there are actually some differences in terms of how you have to deal with these systems?

There are probably some differences. You can probably talk more to the specifics around this, but I think each layer is going to end up being somewhat qualitatively different. But you need to be able to understand the protein interactions in order to be able to understand how cells work. You can't just go straight to cells without understanding the protein modeling.

Mm-hmm.

Then if you're trying to understand something like the way the immune system works, or a bunch of cells interacting together, it's tough to do that without first understanding cells. You might be able to, at a very high level of abstraction, simulate a system, but if you really want to understand how it's going to work, you want to build the simulations at each level hierarchically. That's basically the approach that we're going through, starting with the building blocks and the protein.

I think there's going to be different types of data that you want to collect for each. The modeling techniques, I think we'll see. That'll all keep advancing across the board. But I do think that a big part of the strategy is this view that you need to build it up hierarchically.

One of the things that's unique about us in this space is that we were very intentional that the AI efforts and the wet-lab efforts were a single effort. We've done a lot of work to bring them together. The really neat thing that we can do is try to pull and gather data that helps us connect across the hierarchy. You can look at—

Spatial transcriptomics within a cell lets us look at where it's localizing. We can look at translucent zebrafish and look at development across different cells and when the brain develops. We have sensors that allow us to look at cell-cell communication through different molecules. We can be strategic about the types of experiments and data we want to collect that help us bridge across these, making it so that there's some connective tissue that helps drive the modeling—the modeling magic that happens.

Yeah, the reason I asked the question, by the way, is that I used to be a biologist. I have a PhD in biology, and I worked—

Oh.

—in wet labs for almost a decade and everything else.

Are you looking for a job?

We can talk about that later.

It's not a no.

At this point in my career.

I'm hearing—

It's fine.

I love my aggressive recruiter.

Yeah.

I'm like Danny Glover in Lethal Weapon. I'm almost at retirement.

But I think one of the things that was always lacking was this integrative nature across the different layers of biology. The developmental biologists would work on their own, and the molecular biologists would be doing different experiments, and so that's why I was curious about—

Totally.

Typically, there's a reductionist view of biology, and there's a systems view, and those people didn't really work together deeply. One of the exciting things about what you're doing, actually, is how you're bridging that. That was the basis for the question as well.

Yeah, and if I could add something there, I think we're in the age of this kind of information theory in biology. There are levels of complexity and hierarchy in biology, and each level is made up of and constituted by the lower levels. If you want to have a more complete description and systems that can really generalize and begin to actually answer experimental questions digitally—questions that you could ask in the lab—you need to have the right basis for modeling at every level.

I think what's really unique about what we can do is, as Priscilla and Mark were saying, really build information at each of these different layers, collect them, collect those connection points, but also really do it at the scale that will reveal that underlying information architecture. That's going to be really critical to actually be able to build digital representations that can answer new experimental questions.

One of the things that inspires me most about this effort is really what Priscilla said: There's so much we actually don't understand about biology, and what if we could? I think that's actually very different from lots of other incredibly interesting and useful AI problems we attack. We're trying to replicate human behavior, and a lot of that data is on the internet or captured. Without pretending to understand all human behavior, you can predict a lot of it.

I thought one of the most interesting things in your release was the mechanistic interpretability stuff you alluded to: Can we actually extract new knowledge from what the model believes is happening? Can you talk a little bit about that?

I'm really excited about that. In mechanistic interpretability, traditionally, it's been applied to large language models with the goal of understanding what the representation space of a large language model is, how it computes things, and whether that really connects to what we understand about our intuitive understanding of the world. There's a really rich toolkit that has been developed to start to be able to ask those questions.

What does that mean for biology? One of the classes of models that we train are these protein language models. They're trained on the sequences of proteins. Anything they learn about biology is emergent, and we've seen that they can learn things like biological structure and biological function. That's just emergent from this token-prediction training task.

As we think about mechanistic interpretability in those models, we're really seeing the unknown because the models have been trained on billions of protein sequences. They've been trained on both known and unknown biology, and yet they're developing these representations that start to capture things that we can really see correspond to that reductionist picture of biology that's been built up over the centuries.

You can start to connect the dots between proteins where we really don't know anything about them and proteins where we do know something, because there's that underlying structural grammar linking them in the representation space of the model.

At the extreme, it could be that we're going to understand systems in the body that we didn't before, or the mechanism of action for a new treatment, because we can ask the model and interrogate that representation.

That's right.

Yeah.

The hope is that you really learn the underlying basis for how it's making the predictions, and so you open up the black box and can actually understand the biology that the model is representing.

So, asking for a friend: You guys all believe in venture-backed companies as a way to have an impact on the world. Was it collecting data on zebrafish, or the span of the data, or the wet-lab work, or just the scale? What makes this a better fit for this big nonprofit ecosystem effort versus a venture-backed company?

3. Open Source Spreads Science

Well, I think we just want to give tools to the whole scientific community. I think in order to have the biggest impact, it's not actually clear that we couldn't run it as a business if we wanted to.

I just think that we'll have a bigger impact by getting this in more scientists' hands quicker, by doing it as open-source projects instead. So, yeah, I think that's the approach. But I don't know. It's an interesting question.

Obviously, you were doing it as a for-profit company, a bunch of the modeling before. Then you run into certain issues. You have to raise a large amount of money in order to build the compute clusters. I think in a lot of ways the data is actually even more of a constraint.

Because if you look at the scale of these models compared to language models, they're smaller, but they're smaller because the amount of data is less. In order to get the data, it's not just like there's some factory somewhere that you can pay to produce the data. You actually need to invent novel scientific approaches to be able to do, for example, the type of cellular engineering we're doing in New York or the types of devices in Chicago.

That's why, when we're talking about this concept of frontier biology and frontier AI, the frontier biology is that you need to do real science to advance different biological methods in order to be able to observe the things that create the data that go into the model. It's not just an off-the-shelf thing that you can create.

Now, that's a pretty big effort. I don't know that there are that many things like that that are done as biotechs. I think it's just the scale of the ambition of what we're doing, the time horizon over which we're committed to doing it, and the scale of capital required.

I think part of the theory is that if you're building tools that are this complicated, you want to have a 10- to 15-year time horizon on building out these efforts. I guess there's no rule that says you couldn't do it as an incredibly well-funded startup, but I think that this just made more sense.

It also simplifies things strategically by not having to think about how you're going to make money with the different things. We want to get the models in people's hands. We release them as open source, and I think that's a very valuable thing to do.

Again, the theory isn't that we're going to cure the diseases. We're not. It's that we want to help accelerate the pace of progress for the whole scientific field.

As the person least experienced with making money here, I would say that the neutral nonprofit nature of our work actually helps harness more people to enter this effort. To actually achieve the mission of understanding the totality of human biology and to cure, prevent, and manage all disease, you need the entire academic biotech industry to come together and work on this in a unified way.

That's in part because there's a lot of talent out there, and it's not helpful to leave any talent—exclude any talent—from the effort. There's a super-long tail of diseases. There are the common ones, and even the common ones, I think, if you unbundle heart disease, cancer, and neurodegenerative diseases—even if you unbundle dementia or depression—there are many, many, many subcategories that become more and more niche.

That's not even looking at the long, long tail of rare diseases. Those often get orphaned and don't get brought along when we're looking at the most efficient way to impact the lives of many. But if you decentralize the effort and put the tools in many people's hands, you start getting people who are like, "You know what? I am super interested in spinal muscular atrophy, and that's something I care deeply about."

If you put the tools in that person's hands, they're going to be able to make progress. In a way, if you had to focus your efforts and make big bets, you probably wouldn't, because it's just a niche individual disease—a small-group disease—that will, in turn, if we can understand that disease process, help us unlock a lot more knowledge about how the human body works.

Do you have any thoughts or predictions in terms of what disease areas this work will impact first? I know it's very hard to be predictive about these things, but given the nature of the work and the nature of the models, are there areas you're most optimistic about in the short to medium term?

4. Medicine Treats Individuals

That's actually not how I think about it, at least. The way I think about it is that we want to understand how biology works. In the ideal world, you would say, "I understand the genetics of this person."

I want to think about people at the individual level. I want to understand the genetics of this person. I want to understand the risks they have for different illnesses. I want to understand the mechanistic connection between, say, a gene variant, a protein, and a disease process.

Because if you understand that chain, then you can design a protein or design a drug bespoke to them and actually make an intervention. Right now—and I'm sure we've all had experiences being sick—if you have something that's even remotely nonstandard, you go into PubMed, look up a paper, look up the supplement, and then start going through the methods. You're like, "Am I represented in this paper?"

We're just making guesses. We really have no mechanistic understanding. We're saying, "Okay, you're kind of like these people that we studied, and this drug kind of impacts the pathway that we think is implicated. Let's try and see if anything happens." Time passes, and sometimes it works and sometimes it doesn't.

My goal is to be able to treat the individual as an individual, understand the mechanisms, and be able to intervene. There are different diseases that are at different stages of filling out that whole throughline.

For some diseases, you just want to understand which gene variants actually cause disease and which don't. That in itself can be super empowering to patients. Beyond that, there are some diseases where we understand the chain but just can't intervene and change a specific protein function.

That's super exciting too. If we could design a protein to actually change the physiology, then we could actually cure someone. But to me, that is just as exciting as contributing to our understanding of how someone gets sick in the first place.

Yeah, no, that's a very exciting vision, because you're basically saying you can bring generalizable tools to provide very personalized things for each individual person.

Yes.

And that's the power of the approach: You have these big models that you build that can then apply anywhere. I know that you mentioned earlier that you were going to try and cure, prevent all diseases within 100 years, and you mentioned that it could actually be sooner now, given all the advances in AI. Do you have some thought of when you think we'll be closer to that goal?

I mean, I'm optimistic it'll be sooner. I think the thing that's complicated is that it's a dynamic system, right? If you fix something, there will obviously be future things that you need to work on. So I don't think that the current set of things that we're aware of are going to be the only things that need to get worked out.

But I think the progress with AI is obviously very exciting on this. The other thing that I'd say, just adding to what you were saying a second ago, is that we really look at more systems than specific diseases.

For example, one area that seems really important to understand is inflammation. We talked about this a bunch. This is a big focus of the Chicago Biohub. There's a lot of data on that, and it seems quite clear that it's connected to a bunch of different diseases.

Rather than studying the specific diseases, we think that by trying to understand inflammation more broadly, that will make it so that other companies that can then use these tools can work on specific therapies. Another example is the immune system, which I think is a very good case to study for some of the work that we're doing in cellular engineering, when we ladder up from proteins to cells to whole dynamic systems within the body.

I think that one makes sense. It's sort of privileged: The cells can travel around through the body and all that. Obviously, that has a big part in addressing different diseases. How do you make the immune system function better?

But exactly how you connect that last mile, I think, is going to be more something that biotech or other academics individually studying things will be better suited to do. This is how we think about building out the toolset that helps accelerate all these other folks.

Whether the timeline is 10 years or 100—hopefully, less than 100 now—I think it's useful for maybe your average doctor or patient, human being—everybody's a patient—to think about what's externally visible in the progress here. You worked with patients for a long time at UCSF. What should doctors look out for? What should people look out for if you're actually accelerating progress?

This is the part: I'm super excited about the progress, especially with this launch that Alex and his team have put forward.

And I think it's very clear that science is going to start moving pretty quickly.

Mm-hmm.

I think the thing that's less clear to me is exactly how we translate to the clinic and what that looks like. What has to change is actually the way we do clinical research. My hope is that we're really shortening the distance between bench research and patient impact.

There are a lot of steps there that we need people who actually take care of patients to think creatively and think about how to deploy safely. That's a gap where we have some work to do. We partner with Jennifer Doudna on our CRISPR Cures program at UCSF, so we're dipping our toe in understanding how the deployment of research needs to change, given how quickly research will be progressing. That one is, I think, still shaping up.

Mm-hmm.

Maybe I could say something about our most recent launch.

Yeah.

I think it also illustrates some of this.

Oh, yeah, please. We should ask you explicitly about it.

5. ESMfold Designs New Proteins

Yeah. I guess it was just a week ago, around now, that we announced the new ESMfold. This is basically an open system for scientific discovery in protein biology. It's a world model of protein biology that's been trained. It's a language-model-based system, so it's been trained on billions of protein sequences, learns these emergent representations of protein biology, and then we can use it to make predictions of atomic-resolution protein structure.

And it's really fast. It's blazing fast. It's illustrating this Pareto-optimal frontier of speed and accuracy in structure prediction. This allows us to characterize really vast stretches of the protein universe. We folded over 1.1 billion proteins, predicted their structures, and identified features connecting all of them through mechanistic interpretability.

But I think the thing that I thought was most exciting about this model is that it's this really general model of protein biology. You can use it as a world model. You can actually start to search the space of the world model to design new proteins. It's really hitting state-of-the-art across pretty much every structure-prediction benchmark, and especially on protein-protein interactions and protein-antibody interactions, which is really critical for therapeutic design.

What we found is that you can now use the model to design proteins and single-chain antibodies. You can do all of this digitally and then, in a small number of experimental trials—basically a 96-well plate—select from hundreds of thousands of trajectories digitally, synthesize 96 proteins, test them in the lab in a really short, easy experimental cycle, and find nanomolar binders there. That's really the level for therapeutic activity.

It's really showing that you can have these general-purpose models. We didn't design a model for antibodies. We didn't design a model to be able to bind one particular target. We just designed a model that could understand proteins, and you get protein design as an emergent property.

I also think it illustrates the power of open science and open source, because we released this as basically an open discovery engine, so really anyone can build on it. It takes these really intensive laboratory experiments, where you have to screen through hundreds of thousands or millions of antibodies in high-throughput screens in the lab, and you can really just spin up an instance, compute, and now be able to generate antibodies.

You should say more about how we took that data. When we looked at an antibody screen, we validated it. We looked at PDL in cells, and then we looked at it under cryo-EM, and how all of that complemented and validated what you were seeing in the models.

That's right. I think it's really critical to actually go and characterize these molecules in the lab. We have a structural biology center here. We have incredibly powerful cryo-EM microscopes, so we're really able to look at these proteins biophysically and functionally. We design proteins for several therapeutically relevant targets, and we're able to confirm their function in cellular assays.

It's delightful when it works the way it's supposed to.

Mm-hmm. Yeah, it's very amazing.

We're able to look at the structure also. You can see atomic resolution at the binding interfaces.

Yeah.

Mm-hmm. Correct. I know a lot of your work is really focused on basic research and building out the fundamentals. If I look at actual translation into drugs or drug development, often a clinical trial will be 15 years. It'll cost $1.5 billion. About $50 million of that is often the molecule and preclinical work, and that's a few years of work. The other $1.45 billion and decade-plus are actually on the drug-development side of it.

A lot of that seems to be gated on some regulatory issues. Some of it is recruitment. It's a variety of things, but a lot of it also has to do with the failure of drugs in trials around things like absorption or toxicity. Have you considered tackling that other chain of molecular design and thinking, or is the primary focus more on the basic biology and the initial molecules?

At least my hope in building this comprehensive model of how cells work is actually also being able to predict off-target effects. I think you can do some of that with biological models. Right now, some of the off-target effects are simply things we didn't know, like that your kidney cell also expressed this receptor. Then, when we test it in humans, we see it happening, and we see renal toxicity.

If you have a single-cell atlas that looks at all the different cell types, some of which actually were not predicted before we modeled them, you can start looking at which cells do have receptors for the target you thought you were exclusively targeting and be able to predict some of these downstream effects before we get into human trials. That's one of the more exciting applications of a transcriptomic model: to understand how the different cells will react when you intervene and do something.

But when you think about delivery mechanisms and patient care, that's where you have to start being creative about what disease you want to cure first. There are certain diseases that will be easier to deliver a therapeutic to, or where the risk-reward makes more sense.

We were all inspired by Baby KJ, I think last year now, when the team at CHOP was able to deliver a CRISPR therapeutic to edit a mutation he had—one that would have inevitably led to significant neurotoxicity and altered his life. We were able to do that because the disease was very carefully chosen: We needed to target his liver cells, and we could easily deliver a product that would work in his liver. I think that's when creativity and the wherewithal to choose the right applications can help us unlock the first applications.

Mm-hmm.

Something to add to that: you described the conventional drug-development process, and I think these tools have the potential to have a lot of impact on that process. But what's interesting is to really start thinking about the new paradigms that can open up. What does it mean if the barrier to developing a drug, designing a molecule, and getting through all of those stages is so much lower?

You have programmable biology, and you can really start to create a medicine for every individual patient. I think that has enormous implications for how we do drug development and what the future of medicine looks like.

Mm-hmm. Yeah, it'll be an exciting day when the FDA accepts a virtual clinical trial for phase 1 or something.

Mm-hmm.

Or, you know, one that's based on some person's view of that person. Yeah.

Even short of that, thinking about the specific mechanisms where you see this acceleration, I imagine that if people feel like they can predict impact in kidney cells, or have a stronger perspective on tox because they have this broader understanding, they'll be willing to try many more programs, right?

Yeah.

Yeah. The recruitment could also change, and we have this program, RARE. The basic idea is that a lot of people focus on the most common diseases, but there’s this long tail. The economics don’t quite work out for companies to focus on those diseases, but if you can make it so that groups of patients can come together and organize and say, “Hey, we would take an experimental drug on this,” then, because of the cost that you’re talking about—and how that’s a huge amount of the overall cost—if you can flip that, then the economics make a lot more sense. Then, if you can generate something more easily, you can pair it with a group of people.

I think one of the interesting things from science and engineering is that often you can hit your head against the wall on the common problems—in this case, diseases. But a lot of times, you learn a lot more about a system from finding some kind of rare or weird side thing that’s happening.

Edge case. Yeah.

Edge case.

Yeah.

So I don’t know. I think that’s always been an interesting part of this that actually connects pretty well to this, because now you’re going to be able to enable a long tail of new kind of—

Ideas to get tried and enable them to potentially get tested more easily.

Yeah. That’s a really good point on RARE. In our rare disease cohorts, first of all, they’re incredibly inspiring and powerful, but patient groups are self-organizing patient registries, natural history registries, and biobanks. They’re organizing their own clinical trials. There’s a gene therapy that one disease group has moved forward over the course of, I want to say, 3 to 5 years rather than decades. The speed is so fast because the patients themselves have organized the resources that a scientist or a clinician might need. And it’s incredible.

But I think to some degree you’re going to need something like this because there are going to be many more new things that can get created. But that doesn’t mean that, for the general population, you’re not going to want the same level of vetting that we’ve had historically.

Yeah.

But making it so that people who want to be more on the frontier have the ability to do that is, I think, also going to be pretty helpful.

Yeah, letting people opt in to be part of trials, I think, is one of the big shifts that is starting to happen but could really help accelerate biology in general.

All 3 of you have mentioned, at different points, the power of open ecosystems in such a large space. I think some of that logic around open source and the breadth or diversity of data collection that you guys were describing should also apply in the language model world and the multimodal AI world. Do you think that’s right? Does any of the work you’re doing here change how you think about AI and Meta?

I think it’s a similar philosophy overall, and Priscilla was talking about this: a lot of our focus is building tools that empower individuals to do things. That’s a common theme across a lot of the things that I work on, just putting the technology in individuals’ hands. We don’t believe in this very centralized future where there should be a small number of institutions that are basically advancing all this stuff. Our vision is not that there’s going to be some central superintelligence that solves all of science.

I think people are really important, and I think we’ll be more important in the future. Giving people more tools to be more productive is going to be a critical part of any kind of positive future. And that’s how progress has always been made historically, right? It’s not through centralization; it’s through empowering individuals to try things that are somewhat out of the mainstream, that other people didn’t think were good ideas because they thought they were good ideas that had already been done.

So I think that’s very central to the whole ethos. I mean, to some degree, it’s why you create something like social media, right? To give people a voice. I think a lot of the stuff that I care about is in terms of empowering people with individual AI. Open source is one instantiation of it. It’s not the only way to do it. It certainly is one way that you’re basically saying, “We’re going to take this technology and put it in everyone’s hands.”

In terms of science, I think it really makes sense, and we’re deeply committed to open source. There are obviously interesting considerations here that are important, too, because there are a lot of considerations around biosafety and things like that that we’re going to need to balance and think through how to handle. But I think overall, this is very deep in the ethos of the work that we’re doing both at Biohub and probably a theme for a lot of the stuff that I do. We believe that a positive future is one where you build a technology as a tool, you put it in individuals’ hands, and that’s how society makes progress.

Your mission at Biohub is incredibly ambitious, and yet the AI scientists who work here could also go work in commercial enterprises. How do you think about the talent and how to bring people to Biohub?

Where do I start? I think it’s a very hot market for AI researchers, but part of what that means is that there’s a lot of demand for them. They’re very in demand and can work on the things that they want to work on.

I think this gets back to this point again about frontier AI and frontier biology, right? I think the AI researchers who work here could go work on language models or things at any of the main labs, but those labs don’t have the frontier biology part attached to them. So I think there’s also a very large mission component to this: there’s an ability to do this unique work here that you just can’t really do at the other places. If that’s what your focus is, then I don’t actually think there’s any other organization in the world that’s doing both frontier biology and frontier AI.

Yeah. Why are you here, Alex?

I think it’s really simple. Our mission is to cure, prevent, or manage all disease by the end of the century, and I think it’s just such a powerful—

And you say it with a straight face and a less-than-100-year timeline.

Oh, it’s very serious now.

Yeah.

Yeah.

It’s a really powerful mission, and I think scientists are very motivated by that.

Yeah.

Yeah.

Mic drop. That simple. Yes.

It’s something people are deeply motivated by, and I think we’re at this moment in time where that actually seems like something that can be achieved. We’re building a really unique place where we’re tackling that problem, and we have the resources and, I think, the right things to actually really go after that and do that.

Yeah. That resonates with me as somebody who talks to and hires a lot of research scientists. They want to know if you have the data, if you have the tools, if you have the compute, if you have the talent, and then what the mission is. So I actually think that’s super competitive.

The other thing is that you don’t need a very large team, right? I think it’s an interesting thing about the world: people care about different missions, and that’s good. I think that’s part of the whole and part of why building these tools and giving people the ability to explore what they care about, whether it’s across science or just across everything, is such a powerful way to make progress in society: people care about different things.

In order to make progress in AI, you don’t need many hundreds of AI researchers, or thousands or anything like that. I think you can really make progress with a very strong group of a dozen or a couple dozen people. Finding people who care about this mission is not a particularly hard thing. This is a super important thing in the world. So I think that’s—yeah, it’s just a cool thing about the world, that people are obviously drawn to different missions.

6. The Virtual Cell Comes Next

I think the simplest mental models that folks have, even if they’re paying attention to the space, are essentially structure-prediction models for proteins and protein-protein interaction models. Then there’s this one piece, which is fundamental understanding, and then there’s this theory that someday we’re just going to be able to zero-shot things into the clinic with a much better hit rate.

What needs to happen for us to go from ESMFold 2 to this other piece?

Yeah.

Is that feasible?

I think that's a great question. I would say that I'm really optimistic about that. These are problems that, historically, people could spend an entire career working on: How do you figure out how to effectively optimize a drug? How do you get it through preclinical? How do you do the early safety?

I think that when you have a new scientific paradigm, questions that were once hard become simplified through the new paradigm. I'm very optimistic that many of these core problems will be solved in an emergent way through these models.

Mm-hmm.

One great example of that is toxicity. If you can digitally simulate everything and predict where a drug is going to distribute and bind across the human body, you have the beginning of a solution to that problem. I think that once you have these accurate representations at the molecular level, we're going to start to see really rapid progress on a lot of these core problems.

What's the most exciting use or experimentation with the models you've seen in the last week since release?

It's just been great to see it get integrated into all kinds of things. One of the really interesting things we've been seeing is people connecting it with agentic systems to do automated design and automate that whole process.

It's another example of how you can see bringing together agentic and frontier AI with the ability to have a world model for biology, actually reason about biology, and start to automate the entire design process.

How do you decide what the next step in the research agenda is? You have a world model for biology, and then, just to be very coarse here, I could scale it up, I could add more data. Adding data is a nontrivial thing in terms of new methods and domains.

Do you take input from the larger ecosystem about how people are using it and what would make it more useful, or is it really that we understand the next step of structures or coverage that we're looking for?

I think there are 2 things. We have a view on the next big challenge, which I think is the virtual cell.

Mm-hmm.

Really being able to ladder up the hierarchy of biological complexity to the cell.

Sorry, very basic question.

Yeah.

This virtual cell model—what is the input and output I should expect?

Yeah.

Yeah.

I think there are different views on that. But what you ultimately want is a system that can really model each of the levels of complexity: the proteomic layer, the genetic layer, the transcriptomic layer, and connect that to the phenotype.

You need enough generality so that you can ask the model questions about a new intervention in a context that it hasn't been trained on and get an answer from it. The gap we need to close as a field is being able to really make those predictions that can generalize. That's going to require enormous effort to generate data.

Yeah.

And then, in terms of what you decide to do next, I think this is a pretty normal process of constraint management, right? Every lab and every field across the world probably feels compute-constrained. I think that's probably true here too.

There are always questions: Should we double down more on advancing the protein piece? Should we do more of the cellular stuff? Those are ongoing debates in terms of how you sequence that.

Mm-hmm.

Within that, there's also the question of being at the Pareto frontier in terms of how much you want to train the different models. The size of the models is also dependent on the scale of the data that you have, for obvious reasons.

I think some of it is just where you want to be on the curves and the normal constraints. But this is probably the same process that any research organization goes through: You want to go in all these different directions, and you're just trying to constraint-optimize and make enough progress to do world-class work at one thing at a time while planting some seeds that can blossom over the next couple of years as well.

Yeah. This has been the most dynamic period of technology, at least that I've seen over my career.

Uh-huh.

It's so exciting in terms of everything that's happening with AI, and every week there's something new that's changed.

Are you tired or invigorated?

I'm both.

Yeah.

Wired.

I feel like that's how everyone feels.

Yeah.

I feel like everybody's in this—

Yeah.

—manic phase.

Oh, yes.

Mm-hmm.

It's a combination of—

Yeah.

—invigorated and exhausted.

Yeah. It's wonderful. Things are very unpredictable right now. It's really hard to know what's coming. We have almost like early signs of exponentiation on the model side, with agentic flows that we're starting to see in really interesting ways.

Models are starting to help more and more with models, though that's still very early days for that.

If you're thinking back five years from now and you were to define what success was relative to your efforts—

I know things are very dynamic and have changed a lot, but you have this common thread of tooling for the Biohub. You have a common thread of empowering scientists at scale. Looking back 5 years from now, is there a specific thing that you really want to make sure that you've accomplished or achieved, or a primary goal?

I think we have a pretty clear view of this hierarchical set of world models that we want to build around biology, and the other part of that is that we want to do the highest-quality work in the world.

I think we're basically set up to do that between having a world-class AI research team and this collection of Biohub's world-class life sciences research organizations. I think that's fundamentally a setup that no other organization in the world has.

But you can have a lot of great ingredients, and that doesn't guarantee that you succeed. To me, 5 years from now, looking back, I'm sure other labs or efforts will try to produce things that approximate what we're trying to do. I think that we should be able to do something that is meaningfully better and a unique intellectual contribution to the world.

Whenever you do any kind of research, that's what you're trying to do. If we do that, I think we'll all feel very good. I would also expect that at some point we'll start seeing a lot more idea generation from the people using the models.

I have enough faith that that part will materialize that, for me, it's more about making sure that we do world-class work. If we do, the rest will almost take care of itself.

Very last question for you. Snapshot as of mid-2026: What's the biggest update in your own thinking about Biohub or the domain over the last year?

From the last year, you joined in the last year. I think the biggest thing is that we basically rotated toward and formalized Biohub as the main focus of our philanthropy. This has been a very big shift.

But Alex and the team coming in, I think, has been interesting not only because it's a world-class group. You guys have worked together for a while. I think, also—you talked about how much stuff is changing in the field—one thing that's underrated is that this is an extremely talented group of people who also know each other, work well together, and are stable and good.

I think the compounding benefit of people being able to work well in a stable environment over time is underestimated. That's a really important piece.

Prior to Alex leading the effort, the previous leaders of the Biohub were primarily biologists who were interested in technology.

Mm-hmm.

Now I think this is the point where we really flipped that. Obviously, you have a background in biology as well, but you are primarily an AI researcher with a background in AI and biology.

I think that’s a deep reflection on the way that we expect this is going to drive more value in the future. So those are probably the biggest updates in the last year in terms of the work that we’re doing. It’s a new leader—not just the leader, but a team—that I think has been really good.

And then I think on the rest of the industry, it’s on track. It’s kind of this crazy thing because when you have an exponentially growing curve, I think the way an exponential curve feels is that it’s growing so quickly that the emotional feeling is that it can’t possibly keep going. Right? But the nature of an exponential curve is that it doesn’t just keep going; it keeps accelerating. Exponential growth is accelerating.

I think that has all of these emotions and psychology attached to it, but fundamentally, when you look at the curve in the industry, the fundamental thing is that it is on track. It has remained on that curve, which I think has all these very profound implications for all of these domains. But certainly, it validates and makes one feel very good about making a very big investment in the things that will play out if you stay on that track, and it seems like we are. So that, I think, is very good news.

I think the most important aspect of what you’re doing there is that you’re actually closing the loop with the actual biology.

Mm-hmm.

With code and research, they’re closed-loop systems, and so they’re very fast to iterate. This is an open-loop system, so you’re closing a loop, and that’s really crucial to progress.

Yeah.

Yeah.

For me, one of the biggest changes with the strategy we’re driving now, with Alex at the helm, is that before, we had amazing teams moving generally in the same direction and understanding the potential collaborations and interconnectedness of our work. But now we are arms-linked, moving together—

It feels very directed.

Yeah.

—with a singular goal. It’s very directed, and it’s very exciting. It’s a little bit scary, but it’s truly a team playing off each other and trying to make progress towards this goal. That has taken a lot of work, but also the maturity of our teams, with their work being at a level of maturation where it actually does make sense to interlock.

Amazing. Well, to teams being on the curve, thank you guys for doing this.

Thank you for joining us.

Yeah.

Thank you.

Thank you.

Find us on Twitter at nopriorspod. Subscribe to our YouTube channel if you wanna see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no-priors.com.