Raw Reads: Biomni Press Conference Transcript
A refined transcript containing highlights from a press conference for the Science paper introducing Biomni, from Jure Leskovec's Stanford University lab.
Last week, Stanford University computer scientist Jure Leskovec and former underlings Kexin Huang and Jerry Qu formally introduced Biomni in the pages of Science. Biomni is an agentic AI toolkit designed to help biologists plan and execute complex research tasks, proof-of-concept that with some extra handling, large language models can be further engineered to the exact specifications of life science researchers.
Prior to publication, I attended a virtual press conference with the trio, who have cofounded Phylo to commercialize Biomni. I’ve already spoken about the tool on the podcast and am working on a reported feature, however, I simply can’t include it all, even if I think there is value in it.
Below is an edited transcript of that press conference, highlighting their most informative and thoughtful answers to questions posed by us reporters.
This is the first installment of Raw Reads, a new feature where I share a deeper look at the conversations behind the stories on Ion Genomics.
AAAS (American Association for the Advancement of Science) Moderator
We have Jure Leskovec, professor of computer science at Stanford University. We have Kexin Huang, who is a PhD student in computer science at the time of the work, and has since cofounded Phylo, which works on Biomni Lab, an AI (artificial intelligence) scientist. And then we have Yuanhao Qu, who was a PhD student in cancer biology at Stanford School of Medicine at the time of the work, and has since cofounded Phylo, which works on the Biomni Lab AI scientist. He prefers to go by Jerry.
Jure Leskovec
Great, hi everyone. Thank you for joining us today, and for your interest in our work. I want to start with a simple observation, which motivates our work, is that biomedical research has never had more raw materials, data to work with, and we have enormous data sets, thousands of specialized tools, millions of papers published, and all that. Yet the discovery is getting kind of slower and slower, more and more expensive, and it’s not faster, and in developing Biomni, which is an AI-based co-scientist, our insight was that kind of the bottleneck isn’t ideas, isn’t the data, it’s the fragmentation, the training, the specialized tools that slow down the process, so a huge amount of biomedical data sits unanalyzed, sophisticated analysis that could be done are kind of never done, connections between data are sitting there in the literature, but they are never explicitly being made and harnessed for discovery.
So we started this research by saying, you know, what if every biomedical scientist had a tireless expert collaborator, you know, one that could actually go read literature, one that could actually go run the analysis, understand the data, design the experiments across any, any area of biology, and this is this is what this paper is about, and this is the system that we built, called Biomni, which is a general-purpose biomedical AI agent.
I want to stress here the general purpose, because this is what makes the system different, right? Most AI models, systems in biology today are very specialized, for you know, for one thing, for one specific task, for one specific data type, for one specific prediction, they are very powerful, but they are kind of locked into the single problem, and Biomni is not — it’s a, it’s an agent that kind of sits on top of these specialized, specialized tools, and can execute long-range multi-step tasks across the whole landscape, landscape of biomedicine, from genetics, genomics, immunology, pharmacology, microbiology, and so on, and it doesn’t need to be reprogrammed for each new task, so the way this operates is that the agent collaborates together with the human and gets a request in plain English, [and] the Biomni agent can then plan out and carry out the analysis multi-step autonomously.
During this procedure, there is always an opportunity for the scientist and the agent to converse, to discuss, to ask clarifying questions, to give direction, and so on, and it can take large data sets, from wearable devices to all kinds of omics data, and so on, dig through massive single-cell data sets, and then execute on a given task, and what is interesting is that the agent is able to basically be a companion, a co-scientist to the human scientist, execute the tasks, the human scientist can simply talk to it and can correct it, and the benefit here is that it’s also all verifiable because the agent — the entire reasoning trace of the agent, the entire execution trace of the agent — is actually there, transparent for the scientist.
So, what is interesting here is, I would say, three aspects. One is the generalist nature of this system, that it’s not a point solution, but it plans, writes, and runs its own tools, own analysis using real scientific trusted tools, real scientific trusted databases, and data. So that’s the first important thing. The second thing is that we have actually validated the system and built all kinds of safeguards that produce results that are validated in the wet lab on real data sets, so this system is really validated, not just kind of benchmarked, and then the third important aspect is that this is open — it’s an open-source, freely available system with an easy-to-use, no-code web interface, so any biologist can use it without writing code. We have released it openly with safeguards because we think this is the way to make technology safe, useful for everyone to use.
Kexin Huang
So, I will describe briefly how Biomni actually works under the hood. So, to really build an AI agent that can operate across many subfields of biomedicine, we had to solve two problems, right?
So, first, the agent needs to know what kind of tools, software, databases are even possible in biology — we need to map out the space of actions — and second, it needs to know how to combine them to solve a real problem. So, for the first problem, mapping this action space, we actually also used another AI itself. We built what we call an action discovery agent that basically identified and read through 2,500 recent research papers spanning 25 different subfields of biology, and from each paper it extracted the essential ingredients of doing that research — the tasks, the software tools, the databases — and then we had human experts verify and implement them.
So the result is a unified workspace or environment, what we call BiomniE — E stands for environment — so it brings together 150 specialist biomedical tools, over 100 standard software packages, and 59 major databases, all in one place, one environment where an AI agent can actually use them.
And the second piece is the reasoning system itself, or reasoning agent, which we call BiomniA — A stands for agent — so when you give it a request, three things happen. First, it retrieves just the handful of tools, databases relevant to the specific question, out of the hundreds that are available in the BiomniE environment, because only a subset of these are relevant to the question at hand.
Second — and this is a key design choice — it writes and runs computer code as its way of taking actions, because computer code is very flexible to orchestrate different tools, databases, and software in an intelligent fashion, and we know [large] language models are really good at coding, and this really enables it to string together a very complex workflow, connecting different tools that were never designed to work together.
And lastly, it plans adaptively. It forms an initial plan grounded in biomedical knowledge, then the agent follows each step to perform the actions, and then revises it step by step as results come in, just as a scientist would. We have tested rigorously across a benchmark of more than 400 really serious tasks, and Biomni was very accurate — far more accurate compared to state-of-the-art methods and other leading agents — and then we also benchmarked against human experts annotating single-cell data sets, diagnosing rare disease, finding disease-causing genes; Biomni reached expert-level accuracy while cutting the time from hours to just minutes. So that’s a quick overview of the methods.
Yuanhao Qu (”Jerry”)
I was a PhD student trained in gene editing and also cancer biology, and I have spent years doing a lot of biology work at the bench myself. So, in the Biomni approach, I’m not just like the developer — I’m also a user myself. So, as a wet-lab scientist, I have lots of tedious work every day. One of them is definitely molecular cloning — basically cloning the DNA (deoxyribonucleic acid) to build the construct that we need for experiments.
So, I have been spending several years mastering it, learning how to do it, and I still spend hours doing it on a daily or weekly basis. We wanted to see whether Biomni could help us do tasks like this that we actually do every day, so we gave Biomni one real molecular cloning task, which we would normally do ourselves in our own lab, like inserting a CRISPR (Clustered Regularly Interspaced Short Palindromic Repeats) guide RNA (ribonucleic acid) targeting a human gene — for example, B2M — into one of our gene-editing vectors so Biomni actually did the whole thing from end to end, from designing the guide RNAs to designing the DNA pieces so we could insert them into the sequence, and also wrote out the full protocol we could follow, as well as the primers we could later utilize for sequencing, and also produced the full map of the finished plasmid.
So basically, we then took the whole protocol into the actual lab and ran it exactly as written by the Biomni agent, with no changes, and the next day, after we did the experiment, we found colonies, and then we picked some of the colonies — 12 of them — and sequenced them utilizing Biomni’s own primers, and in the end we found there was a perfect match, meaning the cloning was very successful. To ensure this was not just a one-off, we also ran some realistic cloning tasks that we curated from our daily biology work to really check whether it was actually just a one-off effect, and we found that Biomni performed at a similar level to a senior expert, well beyond a junior trainee, in a fraction of the time.
So for me, Biomni is really changing the way about — like how — its work, work that usually takes me hours, now takes just minutes, so I can really spend my time on the science that actually needs a human.
AAAS Moderator
One thing I wanted to ask was, how is Biomni different from the AI systems that maybe have prompted recent warnings from AI leaders? Can you speak to that?
Leskovec
There is a very important distinction, right? The warnings about safety of these AI models are about frontier intelligence, right — the raw capability, autonomy, and kind of emergent reasoning that is underlying these large language models (LLMs) from frontier labs. Biomni is not that, right?
Biomni is different because it is kind of a layer that sits on top of these models. It’s an orchestration and execution harness that composes existing published tools and public data sets to get research tasks done, so the frontier intelligence is the underlying model, and Biomni is kind of the scaffold around it, or on top of it, and this matters for safety in two ways. The first is that Biomni does not push the capability of frontier models — it doesn’t invent new dangerous knowledge, it makes existing publicly available models, methods, and knowledge easier to run and accessible. And because it’s an orchestration layer, it is exactly where meaningful safeguards can live, right? So human oversight, access control, audit trails are all at the points where actions are actually taken, and the human scientist — and this is crucial — remains at the helm of every step. So, by all means, [Biomni is] a powerful tool, not a decision-maker.
AAAS Moderator
What are the main differentiating features of Biomni compared to other agentic co-scientist platforms, such as … Google DeepMind’s co-scientist? And, you know, the follow-up there would be, why would someone want to choose Biomni over these other options?
Kexin Huang
So, first I want to describe — as mentioned, the Cosmos and co-scientist [platforms]. So, Cosmos is a framework more focused on very long-horizon tasks — often you ask a question and it takes multiple, [up to] 48 hours to get the task done, so on the back end it’s like hundreds of agents working together to get the work done, and on the other hand, co-scientist is very focused on hypothesis generation. So here we can say there are actually two axes, right? The first one is, how compact is the underlying agent — how long is the agent going to run?
Because this is very important for day-to-day use by a scientist, because for the majority of use cases scientists face every day — like running data analysis, doing a simple literature question, or designing a clone — it often does not require like 200 agents working in the back end; you actually want to get it done very efficiently. So, I think that’s one major benefit of Biomni. We designed the agent to make it extremely efficient, often giving you the answer in a few minutes, in contrast to waiting 48 hours.
So, the second dimension is compared to Google’s co-scientist: Google’s co-scientist is very focused on hypothesis generation, which is ideation of scientific ideas, but for us, we are focused on a general-purpose biomedical agent that can work on literature research, data analysis, and designing experiment protocols, so it has all kinds of different use cases. Biomni is more like a day-to-day companion for each individual biologist to work on all kinds of different tasks in a very efficient fashion, so I would say that’s one of the major differentiating factors.
AAAS Moderator
There is an abundance of these general-purpose AI tools aimed at scientists, especially biologists. What is your advice for scientists trying to find the right questions to ask and the right tools for those questions? What’s not suitable at this time point?
Qu
So, first I would say, the Biomni system is definitely designed for scientists, so at this moment Biomni is definitely best treated as a co-scientist for every individual scientist, so I would really recommend that scientists highly interact with Biomni, ask questions, answer some of its clarification questions, and really work with the Biomni system to find the best way to design an experiment and also to analyze a data set.
And I would say, for what’s not suitable at this time point, I wouldn’t necessarily recommend that scientists and users fully delegate the entire scientific research task to Biomni — for example, I would really recommend that scientists really work with Biomni to verify the results, ask clarification questions, and give the whole-picture idea about the research.
AAAS Moderator
Our next question comes from Andrew Han of Ion Genomics. They’re asking, “Are you concerned that Biomni will soak up work that is assigned to junior scientists as part of their training, and maybe even lead to a narrowing of the pipeline if labs decide that they don’t need as many people in these roles?”
Leskovec
Yeah, maybe I can address this one. I think it’s a fair concern, and I would say some of the analysis that junior scientists are doing will get automated, right? But I’d push on what training is really about, right? The value of a PhD, or being a scientist, was never to run the same pipeline for the hundredth time. I think the value is all in judgment, right — knowing what questions to ask, spotting when the results are too good to be true, designing the study, and things like that.
And this is not what Biomni does — this is what a human scientist does — and what Biomni allows you to do is automate pipelines, kind of tedious, fragmented work that can now be done autonomously by the agent, right. So basically, I think the effect of this is that it lets the junior scientist get to the interesting, judgment-heavy part of the work faster, and the mechanical part gets automated. Another interesting aspect is that because Biomni is general, it means that any junior scientist now has access to this seasoned collaborator who is also an expert in, and knows how to use, tools from other disciplines or sub-disciplines.
And then maybe the second thing I would answer is about narrowing the pipeline — my honest view is that history does not suggest that more capability means fewer scientists, right? Usually what happens is the opposite, right? More capability means the value that a given field, let’s say science, can produce increases, which means it gets more valuable, and there is more need for it, right? So mentoring still has to happen — the responsibility to keep training people is on us, not on the software — and I think there will be more need for scientists through this, because there will be more need for the unique human judgment, question-asking, and pointing the systems in the right direction.
Qu
since we launched last May, we have been making tremendous progress, and we have seen scientists across the world really utilizing Biomni a lot and really accelerating their research work — for example, as Jure mentioned, we have tens of thousands of scientists across [the world] utilizing it, and among them we have also noticed that there are some users who have been utilizing the Biomni web platform basically every day within a month, so this is really encouraging for us to see how valuable Biomni is to those individual scientists. And also, during this one year we have been really empowering Biomni with different capabilities and collaborating with other organizations and teams to make it able to, for example, design molecules or optimize a protein sequence, and we have also been pushing it to work with single-cell [sequencing] servers to really allow all different types of single-cell sequencing analysis, and we’ve also tried to really integrate Biomni with some wet-lab automation systems, so Biomni can also not just design a protocol, but also control those lab automation systems to actually perform some of the lab execution. So we have seen so many use cases across how our users are utilizing the platform, and we have kept being amazed by how powerful Biomni is.
AAAS Moderator
Jerry, you said that Biomni can design clones in a manner of minutes at a level comparable to a senior expert. Do you worry that if trainees rely on the system instead of learning the processes themselves, they could miss out on developing a deeper understanding needed to become experts in the field? You may have touched on this a little bit in the previous question, but if you wanted to elaborate on that, that would be great.
Qu
Yeah, definitely happy to talk about this, and as you mentioned before, we really see Biomni as a resource for scientists — it’s great, like, imagine before, you could only access certain expertise during your training process, but right now Biomni is basically bringing expertise from the top scientists across the world within close access to you. Basically, you can interact with Biomni to learn about how to do cloning, which can be much faster than having to figure things out yourself, like identifying mistakes yourself, and you can also work with Biomni to learn more about different biology, new technology, so we don’t necessarily need to recreate the wheel all the time — we don’t need to spend most of our time figuring out how we should learn this or that. I would say it’s actually a perfect mentor, or a resource, for scientists — people can utilize Biomni to do something they couldn’t do before, which is one of the biggest use cases we see. Basically, wet-lab scientists who can now do bioinformatic analysis have been really enjoying and loving utilizing Biomni to learn how to do coding, how to do those data analysis tasks.
Leskovec
Yeah, maybe I can just add on top. I think the risk mentioned here is not new, right? And the way I think of it is, it’s analogous to the invention of the calculator, or a statistics package, or a genome browser, right? And what we need to do is align how we train people — the responsibility is on education and mentors — so that now these tools are here to speed things up, but it doesn’t mean they take the thinking away, and I don’t think calculators are making us weaker scientists, or genome browsers are making us weaker scientists. So, I think Biomni is also — [it’s] not making us weaker scientists; I believe it’s making us stronger scientists, right, because where we need to change is the training, to really let — to say, not “are you able to run this particular protocol,” but “are you able to judge the output? What is the next question we should be asking? Do we understand the process behind this? What’s the research question worth asking?” — and things like that. I think that’s what this type of automation tool actually allows humans to do — in some sense, they free us from tedious work, so we can do things that are uniquely human, better.
AAAS Moderator
If Biomni is free and open to the scientific community, can you explain the business model of Phylo and how it is funded?
Leskovec
Biomni is an open-source research project at Stanford. It was funded through federal grants through the NIH (National Institutes of Health), various sub-institutes at Stanford, and also, when we made it publicly available, not only did we release the code, but we actually provided the compute infrastructure — the tokens, the LLM, [which] were kindly provided to us by Anthropic — so Anthropic provided the LLM tokens, and the rest of the compute infrastructure we provided at Stanford through university funding as well as other organizations. And then about Phylo, Kexin, maybe you can address that.
Kexin Huang
Yes, yeah, so I think, as we know, running this biomedical agent, as already mentioned, actually incurs a lot of cost on the large-model side and on the compute side, and also many other factors, so I think Phylo’s goal is basically to scale with the consumption of tokens and compute — that’s the [business] model — but we also have a general free platform that’s available to academia, because that has always been part of our goal, since we know this is actually costly, and we want to subsidize academia to make sure they can also use the Biomni agent. So we have the open platform that everyone can use.
Leskovec
Maybe I would just like to interject and explain how Biomni operates, right? Biomni, in some sense, you can think of — yes, it’s a website you point your browser to — but whenever you ask a question, this question goes to a large language model, like Claude or [ChatGPT], but then this large language model writes a bunch of computer code that then goes to a dedicated compute cluster where the computer code is executed, data is analyzed, and the output of that is then given back to the large language model, which can then reason over it and decide what new code to write, right.
So to run Biomni, there is a lot of infrastructure that needs to be orchestrated, right. It means there has to be a compute cluster, this compute cluster has to have access to all these specialized bioinformatics tools and databases, and then there’s the orchestrator agent, the large language model, that is now operating the compute cluster to get the work done, and all of this together is called Biomni, right? So just running the system and making sure it’s accessible to a biologist who’s not, let’s say, a computer science expert, is kind of non-trivial.
AAAS Moderator
We see that Anthropic launched Claude Science just a few days ago, which is a general research agent with curated tools built on Claude, which is the model that Biomni depends on. How does that differentiate from Biomni? Is it essentially replicating the same thing? Are there major differences, and what does this really say about the grander ecosystem of these AI computational biological science systems?
Kexin Huang
Yeah, I can take the first part. So, I think overall, the field is definitely converging in the sense that people realize the value of biology agents and how they can tremendously change how scientists work, and I think it also demonstrates a strong validation of our approach, because we started the project almost two to three years ago, when there weren’t too many biology agents, and we released [Biomni] last year, so it’s a great validation for us.
So while the overall idea is pushing AI workspaces for scientists to get work done, there are definitely still many technical differences — Biomni is open source, very available to anyone to install and instantiate locally, in contrast to Claude for Life Sciences, which is more of a closed-source version of it. And then another major difference is that Biomni is also model-agnostic, because Claude for Life Sciences has to use Anthropic’s Claude models, but as Biomni is an independent scientific community and player, we can use any models we want, and we have seen in many benchmarks that different models are good at different things in biolog.
So being an independent player really positions us to route to the best model with the best cost for any given biomedical research task, so that’s also a major difference. But overall, we are excited to have more and more players — when we first started the Biomni project, there weren’t literally many people [working on this], so it’s great to see more and more players, more and more smart people working on the same problems, to really enable scientists to get their work done much more efficiently.
AAAS Moderator
What are its limits? What can’t this platform do?
Kexin Huang
Sounds good. Yeah, I think we were definitely kind of amazed by the capabilities of Biomni, because initially we set out thinking, okay, maybe Biomni can answer some literature research questions, do some analysis, and then we actually realized it’s much more general — it’s automating the intelligence part of the science.
We’ve seen that it’s able to, for example, I think Jerry briefly mentioned, orchestrate lab equipment to perform lab automation, and recently we’ve also demonstrated it can train AI models — because traditionally, training an AI biology model like AlphaFold takes a team of experts in AI and biology, but right now, [Biomni,] because it knows how to write code in the right environment, with the right infrastructure, it can also start optimizing models like AlphaFold on your proprietary data. So we can say the use cases seem unlimited, but we do see there are also constraints, right? For example, physical constraints are definitely one big aspect — we know in biology, physical experiments are one big part of biomedical research, and right now we’re able to show some early sparks.
Our Biomni agent can connect to some experiment via an API (Application Programming Interface) or connect to a liquid-handling machine to get some basic experiments performed at the bench — but I would say a large chunk of physical experiments are still not something the Biomni agent can do yet. The main bottleneck is actually on the physical side — how do we communicate physical experiments almost like a programming language for the agent to interact with — but we are pretty optimistic that as more and more smart players work on this problem, in the future it can become more of an end-to-end, closed loop between the dry lab, wet lab, and everything across biomedical research.