Interview — 2026-08-20

There's a— I'm holding a line. Hi, what is your name? Okay. So I try to find an application so I can have a product that can address the problem. Something that kind of gives you explanation but still has an Yeah, so I wanted to talk to people to see what they think of the images. Yeah, and then analyze that. Okay. Yeah, I think the idea of having a model plus also a wet lab is pretty hot now. Like everybody talks about the need. It's very powerful. Yeah. Okay. Yeah, uh, it might be easy there. Income is easy. Yeah. And then start the production. Because travel is no longer mine. However, I am here to support you about tomorrow. Yeah, interesting. Was it hard to adapt though? I was very fortunate. My first job was with like a magazine, which is a well-known and well-established— I had fantastic colleagues who taught me how to do things and I'm really blessed to be a scientist because that means that everything's an experiment, and every failed experiment is something that like I just have on the record. Yeah, I like that. So you left for a PhD, and now you're in sales. Yes. What brings you to this event? Companies that actually do biotech. Yeah. Yeah. You're going to be back by next week? Maybe. Oh, really? Is it a company or a I he and I actually worked together at the biotech that I worked at. Oh, okay. I didn't realize this was called the bio lab. Oh, you knew that, right? Oh, no. So we both worked together at a biotech startup called PerkinElmer that was doing 3D bioprinting. Where were they in the world? In Australia, I guess. I think they had big like many. I was trying to think about whether. Yeah, yeah. But life sciences industry is actually very. Yeah. Why is your boyfriend studying at the same institution? Is it a Cal State or? I think it's a grad school. What's your name, by the way? Okay. Right. You said UCSD. Yeah, yeah. I thought about going to do my PhD at UCSD, but they just didn't have the right research requirements. Okay. Did you drop out of med school or did you just not go anymore? I guess it's like a complicated situation. Okay. Because like, do you know Brown? Do I know what? I'm Olivia. Brown. Brown? There's like a special program at Brown where when you get into Brown, you also get into med school. So I technically was in the med school program. Okay. And then I just dropped out. I think that a lot of people go to grad school because they think that's what they want, because maybe they don't know anything else. And then they get a little bit more experience, and then they're like, maybe this isn't for me. So in some ways it's good to like fail fast, fail early, right? That's what I think. I just wanted to explore a little bit more. My brother is in grad school. One of the things he regrets is like I've never worked full-time in a lab, and they love science, which is great, but it's really a commitment to go get a PhD. It's like a 5, 6, 7, maybe more year commitment of your life. And if it's not for you, it's not for you. Did you enjoy your police career? There were parts of it that were very dark times. Ryan. Okay, I see. I enjoy the product of my PhD becoming— That's a good way to put it. —a much better person. So it shaped you into a better person? I don't know if it shaped me into a better person. It shaped me into a resilient person with many skills, with a self-confidence that I can overcome. And I think that— I did my PhD in bioengineering. So I'm a scientist and I'm an engineer, and, um, like, I don't do science or engineering in my daily life these days, but science and engineering are kind of ways of being. Yeah, definitely. I like your metaphor. It's very nice. Like, you map the things you learn to, like, how you approach things in real life. Yeah. In grad school, I did a lot of public speaking. Okay, okay. A lot of presentations. Yeah. Do you collaborate with doctors? Because I know UCSF has a medical school. So UCSF does have a medical school. Many of my colleagues in grad school at UCSF are physicians and scientists. But I think a couple of my labmates were like either MD-PhDs or like clinicians doing lab rotations of some kind. It's pretty common, yeah. I would say that physician scientists tend to be interested in certain kinds of projects more than other kinds of projects. Ryan, it's nice to meet you. Yeah. Was biology always like what you wanted to do? Ever since I was 16, I wanted to be a bioengineer. 16? Okay. Why 16? Well, that's when I graduated from high school. And then the freshman year of college, you kind of explored, or like? So I graduated from high school at 16, and I was like, I'm gonna do bioengineering. And then I studied abroad in Spain for a year. Oh, that's awesome. And then I came back and I started the bioengineering program, and I was like, you know, maybe I'll be terrible at this, right? And I wasn't terrible, so I just continued. Okay. And bioengineering was like the only thing that I was ever professionally interested in. Are you in healthcare or like— Professionally interested? Yeah. But not personally? Well, obviously. I just said that, you know, if I were free and retired, I'd probably go volunteer for like, um, what did I say, archaeology? That's super. Awesome. Um, thank you for coming everyone. I'm Rohil, um, the one of the founders of Feathermind. It's the SPC company. We make LabVIEW software. Don't worry, someone will come up and talk to you about it. You guys have LabVIEW. Right after the panel is done. But thank you all for showing up tonight. I guess it feels like we have a pretty multidisciplinary audience across like automation scientists, computational folks, and I think it's a very pertinent topic for you guys. The sort of topic of discussion is not, you know, what Will Server is doing next. I know a lot of you guys are here for that. Buy a beer afterwards. But we're gonna be talking about about lab-in-the-loop discovery, and I think that it's a little bit opaque as to what that actually is. It's very buzzy and in vogue right now, but it's not really like a technical swim lane, right? It's more so thematic, and so many actual technical components are folded into it across AI co-sciences, autonomous labs, computation, property prediction, even scientific data warehouses. So we have a pretty disparate group of folks here across automation, the sciences, and, you know, beta partnerships—every part of what constitutes closing the loop. Ideally, they can help us make some sense of what's actually working, what has been deployed, and where humans are still the middleware in between inference and discovery. So. I'll keep these guys to introduce themselves. Hi, um, I'm Sichu. My first name is actually very hard to say correctly, so I go by Chu in Chewbacca. I'm the Chief Discovery Officer at Zehra, and I've been here for about 2 and a half years when it was still in stealth mode, and I get to lead the high-throughput biology group. We think about assays that are high-throughput for data generation, both for for team design models and for virtual self-education models. Also get to get a chance to oversee automation and we're thinking about how to apply that both for scaling up and scaling out the workflows in the company. Before Trizera, I was leading discovery platform at InCitro, another AI-focused company. Before that, I was at Barely where I was one of the first generation coming in from Google X. Great to meet everyone and I look forward to the discussion. Thanks for having us here. My name is Will Server. So I'm going to start at the beginning. My background is physics and astrophysics. I ended up by luck being the first employee of a company called Synergy, watching that grow up to a thousand people and go public and build a really cool platform. I was the head of automation there. Almost split out of company, and then Ginkgo bought Synergy for the automation tech we built, which was very neat. Spent a couple years. Deploying that internally, and then convinced my boss Jason that we would make a business out of it. So I became general manager of the automation business unit, ran that for a little under two years, did about seventy million in bookings, which I felt very good about for a brand new business. And then about three months ago, I left basically because Jason was now very focused on the automation, and we had a good pipeline and good technical direction. So It felt like the first time I kind of could leave and things would keep going in a good direction. And so I've been thinking about what to do since then, and I'm at the point now of incorporating a company, probably this week. So I'm technically unemployed right at the moment, but very soon we'll be doing— I'll be very busy. Hi, I'm Kavita. Thank you so much to Jack and Rohan for having us here. I lead the product management group in BioHealth, and my day job tends to be sitting at the intersection of our foundation models and thinking through how we validate them for important biological questions that we care about, both on the novel biology side as well as biology that relates to drug discovery. And then ultimately, like, traversing the arc of scientific impact and through how we deploy these models in amazing AI scientist platforms that exist today, like Claude, GemNet for Science, etc. Prior to Biohub, I worked at Google, and I worked across the spectrum of large language models as it relates to scientific discovery and clinical diagnosis. I had a fun ride there, kind of riding the LLM wave where We were the first medical large language model. We launched Med-PaLM, and everything else is kind of history since then. I think the pace of innovation is just insane, and I feel like closing the loop is so critical in advancing the capability here. So I'm super excited about this conversation. Yeah, thanks for being here. So I'm in sales. Hi, my name is Jordan Poole. I am the head of partnerships and strategic deployments for the Science of Living project. I am a scientist by training, geneticist, biologist. The questions in those fields that have always fascinated me are the application— development and applications of different forms of engineering. So that's more or less what I've done over my time. And then the topic, what I am responsible for, what I think about a lot, So, one of the questions that we get a lot is, given the capabilities that are growing at a dizzying pace in the models, what are the areas of science that we believe we can have an impact on? And as we think about those areas, who are the scientists, organizations, companies that we should be working with to close the loop? Right, so we have increasingly fun capabilities, but at least in biology as a science, you need to go do an experiment. That is validation, and usually that requires getting out of the plot and getting into the lab. And so it's fun to think about that and chat with these folks about it. It's great to be here. Cool. Well, why don't we get started? I feel like the first thing that we should talk about here is the component of the whole space that is like sort of predated a lot of buzz in computation and, you know, AI code scientists, which is, as you said, you know, Just going out into the lab and doing experimentation. So automation has been, you know, around well before all of this computation stuff. But I think the way that automation has been sold to scientists and sold and described what it's doing to science— many objectives, right? People describe automation as solving reproducibility, throughput, freeing humans from manual work. But to me, all these objectives seem somewhat competing. in an actual production deployment, and they feel like they each carry their own set of design constraints. So I'm curious across the panel, whether it's recently or a few years ago or whether it's in your current roles, where do you guys really have conviction in automation to solve any of the axes of friction in science, and where's a platform that you've seen it deployed really well? I mean, So, yeah, I think that people do assign a lot of sort of value to automation, or imagine it will bring a lot of value, and it often does. And that could be some of the dimensions you mentioned. It can also be things like safety and the like. Maybe I'll offer up an opinion on what we should be doing that for, which is maybe not the most common value assigned to it. So reproducibility in my mind is a sort of a prerequisite. Like, if you don't have reproducibility, you probably should stop the presses and stop doing your science, whether it's on automation or manual. So I don't think of that necessarily as an outcome of automation. I do think quality is a critical thing that automation should be delivering, where quality is really about the information contained in each experiment. So it's not about maximizing the number of, you know, the number of experiments you're doing, but it's increasing the information density of those experiments. So automation can get you there through like more precision, more reliability, more process adherence. It also gets you there through something humans really struggle to generate, which is like a proper tool trace. Like it really does record and emit information about absolutely everything that is happening. And I think that combined with the scientific data is of incredible value if you're trying to train models And the like. So that's that's really about you know sort of the information contained in each experiment. The other thing that I think is neglected, which is really one of the most important things you could advance for a pharma company or over the life, is cycle times. So if you can get your cycle times down, you know very rarely is your experiment a single attempt at something. Usually you're doing an experiment, learning from it, and doing it again. So you know I would rather instead of doing twice as many experiments. Get that cycle time down from 3 days to 1. And then the combination of these 2 things— quality, you know, that information per experiment, plus reduction of cycle time— really what you've done, if you can improve those, is to improve the rate of discovery, the rate of learning. So I think everyone building automation systems or buying automation systems should be asking about whether the system they are building accomplishes that. That's really the thing to push on. I want to piggyback on the cycle time. 100% agree with that. So we think about high-throughput data generation for our machine learning colleagues a lot, and machine learning is very data hungry. But I want to say that not all high-throughput needs are necessarily addressed only through automation. There are many ways to do that. One lab assay that we think a lot about is these pooled assays where you can deliver genetic libraries in cells and each cell becomes its own reaction chamber. And as long as you have a way to molecularly barcode the genotype to phenotype linkages, then you can scale up these experiments just in a regular tissue culture flask really, really well. Whether you're trying to study cell biology through CRISPR screening or ChIRP-seq, or you're trying to study antibody binding through yeast and phage display assays, each cell is its own reaction chamber. And so if your goal is simply to generate a high-throughput dataset, then there's ways to achieve that through just good old molecular biology and cell biology. But oftentimes these pooled screening assays have caveats. There are different corners they're cutting. And in many cases, the ground truth assays are still these arrayed assays. And so— and the other drawback is these pooled assays often take a whole month or 2 months to do. So these are large— we often call them pre-training data campaigns, where where in 2 months you can generate petabytes of data, but there's no way to really increase that cycle time. On the other hand, these array plate experiments are— could be a lot faster, could be a lot more nimble. They may not be as high throughput as these large pool campaigns, but if you really are trying to close the loop, give the machine learning engineers a week turnaround, answer their question: is my antibody design actually binding to the target or not? Is my genetic knockdown that the model spits out actually going to give my cell the right phenotype? These are better addressed maybe in these automated assays. And that's where I think automation can do a lot of work. Where, in terms of the question where it's been working well, where it's not been working well, there are parts of the automation workflow that can be scaled up. So let's say protein production. That's a very well-established automation workflow. You can buy workstations. Cells that's built on current technologies to make thousands of proteins a week. But as you solve one bottleneck, you often create other bottlenecks downstream. So all of a sudden, how do we QC all of these proteins on HPLCs? How do we measure their binding affinity on CUT&RUN? Once we get the binding traces out, how do we analyze all of these data at the same rate that we can generate them? These proteins. And unfortunately, many of the downstream instruments, they're not really built for automation. They're built for humans to interact with them. So here's where I think, you know, internally we're thinking about ways to use robotics and software and computer vision to bridge some of these gaps. And I think as a field, hopefully there's more smart engineers that think about hardware, software solutions to solve all of the bottlenecks throughout the workflow so you can actually Increase the throughput overall. I think if we look at the scientific method and the arc of the scientific method, right from hypothesis generation, selection, experiment design, experiment validation, feeding that loop back, I think the most important part is the generation of the hypothesis and selecting the right hypothesis you want to go down the trajectory for. So I think that cannot be automated. Hypothesis selection cannot be automated. Even in terms of experiment design, choosing and picking what your experiment should look like in the lab— does it actually make sense? Does it give you the readout that you're expecting? These things, I think there has to be a human in the loop at every step and almost as a gating factor before the experiment makes it to the lab in order for what you get out of the lab to be meaningful addition to whatever discovery loop you're trying to run. These can run self-sufficiently well once you've selected a hypothesis and you've kind of gone down that trajectory once and you can build a recursively self-improving system, but I think that first pass has to have a human in the loop at every gate is just my personal opinion. I'm curious what you think about this, Will. It feels like a lot of, like, the primitives across both digital and then also robotics in automation land kind of designed with some amount of rigidity in mind, like do the same thing over and over again, run the screen at ultra-high throughput, but very few affordances were made for, like, how sort of dynamic I'm curious whether you're seeing investments in, I guess, more of a modular design, whether it's in work cells or even just like data engineering, and who's doing that well. Yeah, it's, it's a good question. I think you're right to point out a major gap. Most automation systems were really designed for high throughput, low mix. One of the things we tried to I spent the last 13 years trying to make a system that could do high throughput, high mix, and it really requires a lot of wrangling. It's a lot of effort, and I think we made a good product. It is not a cheap product. It still requires expertise to use and the like. I also, you know, even at the individual instrument level, to your point, there's just not a lot of easy control often of these instruments. And shout out to Tenderline for like starting to improve some of those things. I can't help also but think about— there was a blog post that Anthropic put out that was just talking about AI's ability to actually make progress in the life sciences and the inaccessibility of databases and the like to that AI. And they basically had to make a deterministic sort of layer between the AI and the database to actually give it proper access that allow it to go find that data. But I think the same kind of thing is necessary for the instruments to actually enable not just the high-level closed loop that I think people talk about most often, you know, where you're maybe coming up with a hypothesis, testing in the lab, and then you iterate, but actually the smaller closed loops that can occur, you know, with each instrument and the scientists just sitting in front of it for Responding to the errors, responding to what they're seeing in, you know, a spectra or that kind of thing. And it feels like with better control and better access, better APIs, you could start to, you know, create those smaller closed loops in the larger closed loop and really start to make progress. And also to kind of generate the data necessary to make lab work more successful and faster. Maybe just to I should comment, you know, on this one and then also the prior. I do think that there's often historically been this trade-off. As a biologist, I felt this, where, you know, automation enters into the realm of manufacturing, and when you understand something very well, you need to scale it up. And generally that comes at the price of the biology, right? You start to shape or restrain or simplify the biology to meet automation. And I think, you know, now, at this point Will just made, I think that there are, you know, historically there's probably a few reasons for that, which is true of many areas of biology. One is, like, the instrumentation is incredibly complex. There are automation engineers. You need a specific person that knows how to do that engineering to change, you know, whatever configuration. The scientist that is trying to do the experiment has to go find that person as opposed to, like, do it on the fly themselves. And so as we think about how automation can be used not for manufacturing but for discovery, and specifically to Chu's point, things like pooled screens and the beauty of figuring out automation with molecular biology where you get to go into organisms as opposed to having to constrain that biology, I think you start to then also open up doors where agents can start to interact and Navigate that, those systems or processes in an interesting way. And so like the pieces are there, but they definitely need to be put together. I want to just one little anecdote. Before we started to make our own automation stack at Zymergen, I actually went to China and visited factories both producing life science products but also car parts and went to an Apple factory. I also went to the Tesla Gigafactory and I thought, okay, sure, Surely we don't have to reinvent the wheel here. We can learn from these and bring them in. And I remember a telling moment where I was talking to the head of the Tesla Gigafactory, and I said, okay, what software do you use to track samples through your factory and to make sure everything's on track? And his answer was, software? Like, it was such a simple factory in some sense. It was just like parts go from station 1 to station 2, and then at the other end a car appears. That they just— their problems were so fundamentally different from trying to do research, that there really wasn't as much to learn there as I hoped. Yeah, it's a difficult problem to crack. That's actually a good transition. I'm curious, so I think that discovery is pretty sexy. Everyone wants to discover something. But it actually feels to me that manufacturing, pharma manufacturing in particular, is actually a much better target for the integration of like AI tools, agents, and certainly automation. The action space is so much smaller and more quantifiable, right? It's various flavors of quantifying variability, characterizing variability, and then trying to minimize variability. Why are we seeing so much less investment in intelligence in pharma manufacturing? Sales? Is it like we're going So wait, investment in manufacturing or specifically investment in— Specifically applying AI to manufacturing. Okay, so I think there's a few challenges to call out there. One is I think it's probably a very good technical fit, but you do have this tension of processes that by design need to be highly deterministic and regulated with models that are not deterministic. Right, and so you can like pair those two together, but there is some work to figure out how to do that in you know a way that if you're talking about batches or GMP or you know whatever the right regulations are, that is is not trivial. The the other, and I think you probably answered this, is I think there's lots of interest in investment. It's just not usually the thing that people will go and talk to you know. Investors or the press or you know wherever about right discovery is fun. I also would just say like I don't think it's either or. We have a funny relationship with prioritization. You can do a lot pretty quickly, and I think it's absolutely like full cement. Honestly, kind of a thought I had when you guys are doing a lot of your evals. I mean, is a human In a manufacturing context, is doing locating more deterministic or less deterministic than an AI agent? That's a good question. Usually, this is a common question that comes up where, you know, the discussion goes as follows, right? Like, this is a really important process, it needs to be like perfectly precise. Where do the models perform? That is your benchmark. But usually the question to ask is like, got it, where does an expert perform? And that is usually not, you know, 100%. And so, like, oftentimes you have to go actually do that eval too. The tests are being validated with a group of experts who process and set that baseline. Well, I think this is, I guess, like, the next thing to talk about is a part of the loop, which is the human connective tissue. Right now, because of the lack of programmability of my fleet or the inefficiency of data capture, Humans do serve as like the middleware between like my intelligence systems and executing things in the lab. You know, I think over time, the very early days when it was analytical software, automation, and prep, very uncontroversial. Those are things that scientists don't want to do. They're viewed as like rote, repetitive things anyways. Now it's come a little bit further, and it's reasoning over literature at an expert level. And soon it's going to be assay design. I guess in the medium term, from like a scientific perspective, where do we feel like the human scientist and expert still drives alpha in accelerating science? I mean, I can briefly comment on that. I feel like I probably answered the last couple of our first simple questions. I think there's still a very high Yeah. There's a category of human intuition that is incredibly important. The models are getting increasingly good at generating hypotheses or analyzing datasets, and that is very helpful. That work generally takes people a long time, especially those that may or may not have a deep training across multiple fields, and all of a sudden you're looking at the brain and the immune system and they can't keep all these cell types straight or whatever. And so the models are are very good at kind of like helping you walk through that and maybe getting down to like 5 pretty good reasonable questions. But a good trained scientist and like understanding the nuances and the trade-offs and really like interrogating those very critically, being incredibly discerning, I think is still a task that like well-trained and good scientists like need to be doing. And similarly, I think that also comes back to some of the design experiments and like the purple work. I think again the models are very helpful, but I would really discourage someone from just kind of completely offloading that work, right? Like work closely with, but like be firmly in the loop. I think I agree with Joe. I think it depends on the— maybe the scale of the loop you're talking about. If it's on a smaller scale, let's say we're just trying to build a Latin loop for— protein design. Then both the automation data, you know, give you data of the robots in action, as well as the feedback data from the design model. Those are easy to come by, and the loop is relatively fast. I can see, even though today it's not fully automated, I can see a future where that loop becomes entirely closed between robots and the agentic model. If you're talking about drug discovery as a larger loop and involves clinical, preclinical data, clinical data, decision-making at critical junctures— do we go to test in human or not, do we push this into LSR and allocate a lot of resources to make a drug molecule or not— then number one, I think there's a lot less data out there about how to make good clinical stage decisions. Number two, stake is a If you put a drug that's not well developed into humans, that can really do a lot of harm. So I think it depends on the scale of the loop and how much data is available, as well as what the stake is. Okay. I'll maybe do something slightly foolish and I'll give you a hot take on a question you didn't ask. So I would separate the moral from the practical. So on the moral side, I feel like if I think about AI's role in art, I think it's very reasonable to have strong feelings about that one way or another. On the science front, I have no qualms, basically. I feel like if we can do science faster, that is just better for us. And so anything that advances that, I think, is worth pursuing. And so it doesn't worry me about— like, I don't actually think, practically speaking, humans are likely to be removed for a very, very long time, for the reasons already mentioned. But if we I would still snap my fingers and do that because of the value of advancing science. I would also answer on the practical side, like, I think there's a ton of value in humans monitoring what's happening and exploring the weird stuff that happens. You know, if I think about the cosmic microwave background, like, that signal was noise, and they were trying to figure out how to remove that from their radio telescopes, and it was, you know, it was only It's only when they couldn't prove it, they realized that they had actually stumbled across a fundamental truth of the universe. And if you don't have a human with their intuition looking at that, I doubt you'd find that. That's my guess. I think this is honestly kind of an interesting topic to explore because when you think about where, not even AI co-scientists, but just agents in general are most widely adopted, across industries, it's mostly just information compression, right? Taking large amounts of data and telling humans who are driving them what the signal is, like installed invoice. That actually feels like one of the most human-like and intuition-driven types of tasks. But when we think about like the other end, we view like certain types of human intuition as more sensitive or like pertinent, just like designing or executing science or maybe post-processing data. I guess like, where do we feel like, like, you know, over the next 5 years or so, regardless of how good the agents get, we still want a human face to attribute to a particular decision made in discovery, lab, in research, wherever it is? Yeah, I would say it's still like problem selection. I think should be largely human, because even with well-cited publications you have conflicting evidence. And so how much do you rely on extraction across literature and synthesizing this data across published research without a human kind of opining on, you know, what actually to take to the, you know, take down the loop in like why. So I think the question to ask still kind of solely belongs to the human in the loop. And I think data processing can be automated, but I still think data curation should still be human because there's so much of conflicting databases that even just a simple aspect of like PDI, if you look at The databases for protein-protein interaction, for example, you'd see a lot of conflicting data just for that one simple task. So I think it's very important that the human defines the problem statement and narrows down the scope of research, if you will, and the search space. I feel like we're all kind of like centered around the problem selection one. The one thing that I'm a little I'm a little surprised hasn't come out. I think, like, yes, it's true that I think it's important from a scientific rigor perspective. I would also argue it's probably quite important from an alignment perspective. Mm-hmm. And, you know, ensuring, or at least like some amount of oversight there, right? Especially given that so much of what we're doing is dual use, or at least has that possibility from a capability perspective, that like oversight at a minimum, probably like being firmly in the loop there, Yeah, sort of combining those 2 thoughts, I feel like knowing when to stop is actually a thing that I would love to have humans in control of, either because it's a bad idea to continue that research, or maybe, you know, I feel like AI can sort of go very deep down holes and it doesn't know when to actually sort of pull back and go, well, maybe this I think the last question before I let the audience ask a few questions here is around data partnerships. I think that folks on the ML side are starting to see labs as data-producing engines now at this point, and I understand the utility of multi-site or pooled datasets from their perspective because of the heterogeneity, volume. It just makes sense. But it feels like for a particular lab or research team, the actual data produced is very representative and driven by both their scientific strategy and their automation platform, just generally their experimentation platform, I guess. What are people in the space establishing data partnerships, publishing open datasets What are they actually viewing as tradable versus proprietary at this point, and how do they protect their IP? They can take the first crack at that. So part of the data that we work on— part of the model, I should say— that we work on are these so-called virtual cell models. And everyone has a different definition of virtual cell. Our definition is a model that can make counterfactual predictions in unseen perturbation, unseen cellular context. That you haven't measured in the lab, what would a perturbation do that maybe has a desirable effect, making my healthy or sick cell turn towards a healthy phenotype? And so to that end, we did a lot of these high-throughput perturbation experiments in the style of HotDrop-seq. And that was a method that was published in literature a couple years ago, even at the scale that we're doing, but these are very difficult to scale to systematic scanning of the whole human genome. We're talking about growing a full incubator full of cells, each of the cells perturbed with a different genetic knockout, harvesting all of them and generating high-throughput single-cell RNA-seq data. And, you know, that's just $1 million just in raw sequencing consumables per experiment. So when we first began to do this, we spent probably a good year optimizing lab protocols, making sure that it is highly consistent, AI-grade, just because it's very difficult for an academic lab to do those type of repeat tests in every step of the workflow. And we made that protocol available freely with the community, and also the first 2 genomes go up to Terpsichore freely with the community. And we think this field is so new, there's so little data available out there Hopefully setting some standards, making the protocol available will help catalyze— and making some data available will help catalyze the whole field. And the cell lines were not of particular disease alignment with our internal effort, so we don't really see that as a big issue. So maybe that's where we draw the line, you know, generic protocol, we're free to share with the community, but Cell lines that are specific for one particular disease area that we're pursuing internally, that we'll probably hold private. But we firmly believe that open data, open protocol is one way to catalyze the whole field. Cool. Well, I think those are all the questions that I had for the panel. Why don't we kick it over to you guys? Last name? Yeah, but I'm a scientist. I did my PhD at UCSF. Do I just pass her around? That's a great talk. So there are all these new generative protein models coming out, Confusion3D, I've heard. I'm curious, like, are there any ways in silico to rank or evaluate generative de novo proteins other than just validating them one-on-one? That's a, that's a good question. So, you know, if your task is to design a Protein, let's say you're trying to design a binder against a target of interest. I think most labs do in silico filtering in-house. For example, you look at an inverse, so your folding is going from sequence to structure. Design is the inverse problem, right? So when you have design, you could go with a forward folding prediction to see if it's a good model. I should say that that works to some extent in silico. But I think it's a data distribution problem. If you— the target you're trying to design is very close to the training dataset that's available in ADB, you probably have a pretty high accuracy to predict whether your design binder docks well or not. But it's very— if it's very out of distribution from the training data, the clinical accuracy drops. Same challenges in virtual cell, right? If your prediction is very similar to training data, the cell type is very similar to what you have already perturbed in the lab, chances are the virtual cell model will do a better job. But if you're trying to veer out into different cell types that you've never perturbed before, chances are the model will not do that well. I think that's where it's very critical to have integrated wet lab capabilities so that you can falsify a hypothesis made by the model and improve the model recursively. I think people are trying different methods out there, but I, um, I would say, you know, human folding models are still probably the most efficient at high throughput. We are talking mostly about preclinical, but if you really want to solve all diseases and aging within 10 years, we should also talk about clinical and human data. Where do you see that? We don't have enough human data, especially not Agreed. It feels like coming up with new proxies for clinical results. That feels like an area that a lot of effort should go into at this moment. We couldn't agree more. Quite honestly, it's easier to scale data on the in vitro side because you have those— those are renewable resources. You can generate cell lines quite easily. It's a lot harder to generate clinical data for 2 reasons. Number one, clinical samples are just harder to come by, especially for diseases where biopsy is not a routine part of the clinical procedure. Or 2, it's hard to get clinical data that come with well-matched clinical annotation, and those are often what makes the sample, the molecular data, useful. And so I think those are the areas that it's hard for a company to take on that. But I know Biohub has a lot of them, or I should say broadly, there's a lot of nonprofit consortiums that have made great efforts in into curating human samples and analyzing data at scale and sharing community. So I look forward to those efforts as well. Hey, this is going a little bit back to generative protein folding models. A few months ago, like 5 or 6 months ago, a lot of companies like Chai, Boltz, CommandFold mentioned that they basically were able to make nanomolar binders from scratch. It was pretty much all within a month window of time. And so A lot of people have hypothesized that this is just making the models larger, using more protein data, and so on and so forth. How much do you guys— how much have you guys seen proprietary datasets actually improving that performance further versus kind of the classic, you're just scaling these models bigger and emergent behavior doesn't have more bias? Yeah, I can take a stab at that. So we launched our Protein foundation models in the month of May, the DSM family. So we definitely did see scaling laws at play there. So obviously, like there is this thesis: if you feed it the right kind of data and scale compute, you are going to get better with your performance. Having said that, I think the architecture of these models is very critical. So we've kind of built our folding model on top of a representation. Representation model. So our thesis is that learning the grammar and the underlying representations of proteins is very critical in order for you to make structure predictions with high confidence. And so that is the thesis, and we've been kind of seeing scaling laws work really well in terms of actually having some interaction analysis of like, you know, how do these proteins interact and like are we able to predict this with high confidence and do our affinity scores match up when we design these binders. I think there's a lot of like play of generating very specific protein-protein interaction data to feed these models because what it has seen is not sufficient, right? So I think I think the next scale of effort should really be around, like, mostly PPI data so we can actually get a better understanding of how these proteins are dynamic molecular machines and how they interact with each other. Right now what we have is a really solid view of their static structures. So dynamic structures is where I think the investments should go. And also in terms of like function, I think that's a very critical aspect of proteins we want to try to map out, and having interaction analysis data will better inform function as well. Just to— that sounds interesting. You mentioned function, and I got my eyes just lit up after you said that. I 100% agree with what you said. In some way, I guess folks are not surprised by how rapidly this field has developed. We're sitting on a treasure trove that's pretty new that took decades to curate. Rarely in biology do you see such a well— high-quality, well-curated dataset with great metadata. And so I think we're seeing the result of that is rapid innovation on the machine learning model side, and the models are iterating very quickly with increasing performance. Having said that, I think binding is only a part of the equation for making a good drug. For those of you who are interested, I encourage you to check out a recent blog post from Zara on progressible binder. There's a lot more to binding function, as Kavita eloquently pointed out, but also the solubility, including imaging, aggregation, all of these properties. And when you talk to experienced drug hunters in the company, they will tell you all of these matter, and you almost All of the other domains, it's much harder to get scalable data. So there's a lot more work for biologists, for computational engineers altogether. Think about how to generate data for all facets of a good antibody drug so that we can have truly a zero-shot model and give you good drugs. It's job security. I'm going to abuse my position and ask you a question, actually. John, I feel like my impression is the architecture But do you think the scaling law actually comes from the quality of data and basically there being causal data? Yeah, definitely. I feel like we don't know enough and we don't have enough data on, say, IDPs or IDR, for example. So this is something that models routinely struggle with, right? And so it's so important, the data to feed these models. Is curated and to some extent has some validation, like how we deposited structures within the PDB, we should try to resolve as much as possible within the IDR and the IDP space, for example, to be able to train, you know, prepare better pre-training data for these models. In terms of post-training, yes, like everything Chu said kind of makes sense. Solubility is important, and I think closing the loop on each one of the, say, properties almost of the protein, from solubility to stability, etc. So we can build many loops almost to close the loop on how we optimize for those specific properties. But I think the pre-trained data must contain, like, more discovered proteins, if you will. This is for Chu specifically. So everyone has high-developed proteins. Oh, okay. It's kind of scaling with the model size and the data size, but there's this evolutionary factor that makes that structure nice. You don't see scaling laws with chemistry or even really single-cell perturbations. Linear models are hard to get out. What do you think Is the solution here? Is it quantum computing? Is it just a million times more data? A million times bigger models? Great question. On can't comment on chemistry too much, but on single cell virtual cell models, I think there's some controversy about whether there's true scaling law or not, and I think that really comes down to. The data and the task. If the data is profiling data of healthy human tissues and the task is whether the model can classify a T cell from a skin cell, then I would agree you don't need a lot of data to achieve that as actual performance. But we, when we work on the T cell model, we define it slightly differently. As I said earlier, we're hoping the model can give us novel target ideas, learn novel disease mechanisms. And this is where we need to encompass a broad range of biology, different cell types. But on top of that, hopefully we can understand patient-to-patient variability. So different donor backgrounds, different drug perturbation, genetic perturbation, stimulus, time points, readout modality. Only with that can you start to understand how to have a model that to make causal predictions. So in a recent publication that we put out at Cell in March, we showed both the data and the model parameter actually scales quite well, and both of them can beat linear baseline as well as other popular models in the field quite well. And I think that's where we see if we perform perturbation experiments across 7 different systems, there's increasing learning in biology. As you scale the data that you don't achieve if you simply scale cell numbers in the same cell type. And again, if your task is to ask these diverse unseen prediction questions rather than annotating cell types, again you see that you need a lot more data and you need a lot bigger model. So I think scaling model to me is probably real. It just depends on, you know, whether cell number is the relevant metric or how Hello, I have a question about data generation. So in the context of the organizations, how do you think of the data generation strategy that is primarily focused on generating data for training models? So, I think that's a very good question. Yeah, I can comment on like the— I mean, the training data I think I covered, but on the deployment side, I think a lot of the validation around what we're predicting as the function of these proteins could come through validation as our thesis. So we put out So, we have a feature atlas, for example, for 6 billion proteins in the metagenomic universe. Now, we obviously haven't validated the features of these 6 billion proteins, so this must come through, like, actual kind of scientific exploration of, say, a focused cluster of proteins and somebody coming in and saying, yes, this actually is true and these feature characteristics hold up versus not. So we do think a lot of, say, the loop can be closed through actual deployment as well. Yeah, I would only add, I feel like true to your answer also to the previous question, but it feels like you have to think very carefully about the space that you're covering. Assuming you're exploring a fitness landscape, actually getting good coverage In that space, having the right replicate number, having the right sort of process information about how you created the sample to give it context. Like, these things are all critical. And I know this is like sort of a non-answer, but it does feel fundamentally different to produce data for training a model as opposed to giving it to a bioinformatician to, you know, uncover some small fact. Like, those feel like fundamentally different things and require a different set of skills. Maybe to double down on that, I'll just add one thing on this and also touch on a point she made around the single-cell data. You need evals, right? And like generating a whole bunch of data because like gold will be at the end of the rainbow, like maybe, you know, I'm not saying it won't work, but like if you know how to measure the rainbow or what gold looks like, much better. And I think right now in biology, certainly in single-cell biology, the problem is we are using today's models for yesterday's problems. And no one is spending the time, or few people are spending the time, to think about like tomorrow's evals. And are the models unequal climbing against those? And if not, what are the datasets that are necessary? Couldn't agree more. Hey, uh, I'm not exactly sure how to ask this question, but it is fundamentally an alignment question. Um, all of the automation and AI I think AI is going to help us find the right answers in biology in ways that will— I agree with you, I just want to know the right answers— but might remove the step where the humans are actually doing the wet lab biology. And I think that many people in this room really benefited from years of toil in wet labs. The reason why I understand the problems of bioreactor scale-up is because I tuned those PID loops by hand for many years. And, and if we are saying that the intuition of the humans is a critical factor, that intuition came from actually interacting with the pipette and the petri dish by hand. What is changing in academia in for early-stage scientists when, you know, these tools that are really taking things out of the hands of those scientists still need, you know, direction? Does that question make sense? Yeah, and, um, so I work for a company that creates recombinant proteins, which are really valuable for antibody drug discovery, for example. And a lot of people are using AI for antibody drug discovery. So I just came on by. Interesting. Makes perfect sense. I don't have a great answer for you. It's one that I'm asked a lot. I think it is something that we are thinking a lot about, and the world is thinking a lot about. And given the pace and the diversity of tasks that are in question here. I don't know. There's, I think, 2 things that have really come to mind. One is just because a task can be done by a model doesn't mean that either you shouldn't go and validate it yourself, meaning like I can say apply my single-cell data and it'll crank through it and give me a heatmap at the end and a Python notebook. It may or may not be right. I should still go and like at a minimum I'm going to go through that Python notebook and maybe even validate it myself. Net, it will save me a ton of time and hopefully helps me understand what is going on. I think there is like something there that has not been formalized but needs to be formalized and probably needs to evolve in graduate education and like across other areas of science. Number one. Number 2 is, you know, just speaking very personally, There's a lot of things that I do because I enjoy them or I want a connection to a task. I like to garden and grow things. It is much more efficient for me to go to the grocery store, and in some cases those vegetables are in fact better, but I enjoy it. It gives me a connection. Actually, in a lot of other instances, my vegetables are like weird and like more enjoyable. I'm not saying that like science will be like artisanal in that way, but it is simply to say that I think A, A, back to the earlier question, we're a long way from models running autonomously on all these tasks. We're a long way for biology. B, I think we need to be thinking about this training because that toil, and if we're going to be engaging, we need to be learning alongside and with, so staying in the loop. And C, I think it's also hopefully efficiency gains along the way, but also probably some tasks where we'll just remain much better, or it's enjoyable, and all those things are validated from more of a personal point of view. Could we do one more? Right now, I think machine learning is very much in the early days of genetics. Like, you're doing one gene knockout by one gene knockout and trying to understand function. Similarly, we are trying to ask it for one thing and then developing the next model, the next thing. At each step, there is a humongous data collection, and then maybe we solve for one thing well. You are all at premier institutions/companies. Have you thought about from here until, like, making a predictable drug that just works seamlessly in the clinic, what are some MVP human models, and how can you maybe jump the loops of some of those steps where you reach a better answer just by simulation, whereas smaller dataset can help you loop it in. I'm not sure I totally got the question, to be honest. So you're asking about like where smaller datasets can improve specifically in kind of preclinical development, or— Right, instead of thinking like each step of data generation as an isolated point, what can you generate Smaller datasets or generate some datasets that can help you jump a loop, and then you progress instead of linearly. You progress like at a logarithmic scale in this step where right now we are taking baby steps towards simulation of understanding oh whether a drug binds to a protein. But instead of that, you bypass a few steps like because it binds or because interacts in a certain way, you know the pharmacokinetics and dynamics. That could be like. I can take a crack at it. I don't know if this directly answers your question. I think on the— maybe on the meta level, if your question is about how to demonstrate success in the clinic quickly, then I think all of these protein design models are actually taking the right approach, right? If your task is how do we define the right target or not, that will take many years to prove. Yeah, I think that's Preclinically a good point. and eventually clinically efficacy. But I think all of these design companies short-circuit the first part of discovery. Maybe take well-validated targets that were traditionally challenging to drug using old methods of finding the antibody, and they can go to the clinic much faster. Maybe another way to answer your question is if we're— let's say we're talking about building a purpose online, and If our goal is to build an all-knowing biology model, that's going to take a long time. There's a lot of biology to be collected, and we talk about measure modality, time scale, donor, but even there's different scales of biology, right? There's a sequence scale, molecular scale, cell scale, tissue, whole organism. So that's going to take a long time, but it doesn't mean that we couldn't have short-term wins from a model. And so what that Rather than going comprehensive in our data collection campaign, can we be very targeted and intentional in data collection to just answer the predictive question? Back to Kabila's question, have human intuition guide the question selection, data selection, or even have human ask the question but have the model propose what are the data distribution that it needs to improve its own predictive capability? I think Those might be short-term wins where we can cover just a continent on the whole world map of biology, but be able to deliver short-term wins. Cool. Thank you guys for coming out tonight. I appreciate all you guys for doing this. So you guys hang around for a little bit. Thanks. Do you have LinkedIn or something by any chance? So you're not a research scientist at present? That's his card. Oh, you have it? Awesome. I'm assuming you attend a lot of these events. Yeah. Cool, it was really nice meeting you. No, I'm not a research scientist anymore. Interesting. What happened to the switch? Like, why did you go into— I love talking about science. Okay. Okay. That makes sense. Yeah, thank you. But being a bench scientist is very taxing. Sure. Okay. And I had done it for I started doing event science, like full-time event science, in 2013. So I've been doing it full-time for 10 years. And I was like, I got offered a job and I was like, I'm gonna try something else. Okay. What about you? I guess like I dropped out of med school. Okay. So I got into med school, but then I decided to come here to basically help my friend with a company. Okay. He's doing recruiting for different companies, including healthcare companies. He wanted me to be in charge of like recruiting for healthcare. Have you been to many events here? In SF or SPC? Okay. Like, open evidence, like, science-friendly. So I'm like helping. I guess the thesis is that a lot of companies do like recruiting by, like, I guess like spreading like— they just reach out to people that might not be a good fit. But what we want to do is find people that are a good fit for the company and then connect them to them. So it's more of a relationship company than like spreading like resumes or spreading like— yeah, I see that. Yeah, yeah, yeah. I think it's the best label I've ever found. And our job, right? And our job. David? Yeah, uh, we are creating— well, the goal is to create that. I guess we're pretty early, but we're starting by creating tools that How it works is like with every post writer, and then on the candidate side, we make the hard job. So not the robots that have these like natural conversations with the recruiters and stuff, like who they are as a person, like what kind of culture, like I guess like what kind of culture they'll do the best at, and figure out like based on the culture and that's never happened. Okay. So most people, I guess the CEO, like who he's looking for. Oh, I see. Okay. But we found out that like a lot of candidates that come in are really good technically, but they just never passed the coaching exam. That can be a lot of times. Yeah. Yeah. So, so, so I get it. If a coach and I are not compatible, we would like, let's say, send a few like compatible people or some stuff. So I basically Yeah, so pipette is a good example, right? Not even pipette, but let's say for our first application is purification of proteins, I guess, right? Where you might be using magnetic beads or resins, building pipettes, or even centrifugal instruments. But a lot of these are slow and to interdiscipline is a challenge from what we've already experienced ourselves. So we're creating materials that are automation worthy and allow you to do that purification faster and more efficiently. At this location. Yeah. Yeah. So the beauty is our materials are really compatible with both, right? Yes, you're thinking about multiplexing or high throughput. Yes, maybe the automated liquid handlers could be great. But yeah, if you want more dexterity or everything, robotic arms would be great. So we have a material that is sort of compatible with either of those. Oh yeah. Yeah. Yeah. Sorry, we're in Glasgow. And now I'm thinking, so okay, I'm not from here. I'm from Paris. Now I'm here for my master's. Oh yeah, nice, that's awesome. Oh yeah, yeah, yeah. So you're completely based here? No, so even I'm not from here, I'm based in Boston. So yeah, I'm visiting for a few days and yeah, this happened to be perfect. Yeah, I think you're like the perfect example of what somebody Oh, greasy. Yeah, very, very, very, very fresh. So like I said, we are going after proteins first, and within that, with early-stage biotech companies who are developing new protein modalities, and how do we help them generate data faster and more efficiently. Yeah, very true. But you love your current company? No, I've heard very good things about William because he was the founder of Cyphergen and then it got acquired by Ginkgo. And then I've seen, I've heard Jonah speak as well, but I don't know any of them personally. So I talked to the guys from Ginkgo not a long time ago. So I was just asking like, which one asked me like, what do they need to do to become a good scientist? Yeah, yeah. Okay. People in the web app, at least for— I guess you can ask some people from the Academy now. Yeah. Okay, now we should get sending students somewhere else which we paid by ourselves. Yeah. So I think the big key is trying to figure out what kind of people that you want eventually. Yes. So, um, I think, yeah, how do you outsource the work, but still from That's true. Not everybody in the Bay Area. Yeah, in general, most people are on a growth trajectory. I would say so. Oh, this is my first time. So you just moved here a couple months ago? Like 3 months ago. Okay, you're just a baby. Yeah. How about you? Oh, I moved here 20 years ago. So are you working at the university? Oh yeah. So within the living environment, are you like Really? No, no. Oh, for my— I don't have LinkedIn on my phone, but I'm happy to. Yeah, this year the company I work for paid for it. Um, yeah, my PhD took a really long time. Yeah. And, um, so I mean, I only just started attending these events. So do other people that are like seniors— so do other kind of I know that, uh, yeah, they have— I remember they even sell the chicken neck from 45. A lot of people— yep, I can't remember the actual number, but the beef is quite decent. Okay. Yes, I mean, I still, uh, I guess I've had some people describe me as like Korean because I Graduated from college. Yeah, it's only one step into getting used to having to kind of rip out your tail, stuff like that. So it took— so it's not that far to go from here. You do well. Okay. Rip out your tail. I was gonna say, because like, if I was a girl and you're not close, I would think that you would be my favorite. You also feel that way because— so, uh, it's more like— hi, Dominic is hosting me, man. Okay. No worries, no worries. I think he's just so busy that he's not even here. A couple years. I was coming to visit, but I thought he was busy, so no. So we catch up, we should finish up. Yes. I would describe myself like inexperienced scientist, but still like software salesperson. Okay. Yeah, so the idea is, I guess, like, like, right? And I know Saithiba has Right. So the idea is how do I replace these beads and make it not only faster but also more readable because these beads tag randomly, right? And hence it's difficult to display and make it more readable. I thought it was fantastic that somebody who was in the audience actually connected with me on LinkedIn while I was still talking and like chased me out of the room and cut me down. And then over the course of a couple of years, she tried to hire me. While I, while I said no, I would like still have to pursue a PhD. And then eventually it got to a point where there was a job available and I was tired of being in academia. So your previous roles were also sales roles? So my current role is currently my third sales role. So I was at one company for a little bit more than 2 years, and then I was at another company for 4 months. And now I've been at my current role for 6 months. It's very expensive for salespeople. I mean, more than I would have thought a while ago. Okay, I see. Yeah, but I mean, frankly, think about salespeople in fields who have been selling the same products for 25 years. They, they have— they seal it off. I've never been to this location. Yeah, yeah. I guess your bio background really, really helps you. I would say that in life sciences, most of the Our salespeople have PhDs or master's. Yes, because our customers— our customers want to talk to a scientist. They want to talk to somebody who speaks the same language they do, who understands their needs and their struggles. And also the science is quite complex, right? And so if you have a sufficient scientific background, you can look at the information that the I don't know if you want to go into a heading or like maybe try to talk like this. I think you really, really understand. Okay. Yeah, I might have to also try to talk about this part in the opportunity. Yeah, I would say because what I'm doing now is quite applicable. It's still very applicable. Yeah, so it's fantastic. We both consider everyone. Oh yeah, how's the startup feeling, Paris? Quite different. Okay, yeah, not a lot of good crossing. I would say Gustave and also the mainly— I see. I would say the Spanish kind of crossing Paris. Also the global bank, they also organize or clean up the streets. Really? Oh yeah, in Europe they have a beautiful system like So I still try to find application. I don't want to sell it as a model. Yeah, yeah. Because now there's so many of these out there. OpenAI will do the job.