In the early hours of Friday, July 10, a team of researchers dropped the world’s first general-purpose biomedical AI agent, which has the power to autonomously complete complex tasks that would take a team of scientists days, if not weeks, to do.
Biomni, built by a team of researchers at Stanford University, is being called a “co-scientist”, with real-world case studies demonstrating that it can interpret multimodal datatsets, generate experimental protocols and basically do anything required of a scientist working in biomedicine.
“If you think of an agent as a carpenter, a carpenter without tools is just a carpenter who can talk,” says senior author Jure Leskovec, a professor in Stanford’s School of Engineering, explaining what sets this model apart from AI chatbots. “With Biomni, we give the carpenter a set of tools so it can build.”
Biomni Demo Interface
Its description as “general-purpose” undersells its skill set, as it’s less a “jack of all trades” but an expert in quite a lot of them. It can read relevant scientific literature, form hypotheses, select datasets and tools, write code, interpret results and then put forward avenues of further investigation.
“Biomni is able to understand a simple question like, ‘Why are these patients responding differently to the drug?’” explains Kexin Huang, a former doctoral student in Leskovec’s lab who now runs startup Phylo, which has released Biomni. “Then it digs in, doing a lot of the scientific legwork.”
While this tool isn’t necessarily that useful to the general public, it’s extremely exciting for the science community. To date, most models have been highly specialized and limited in scope. Biomni offers robust skills across the fields of genetics, genomics, pharmacology and more.
As detailed in the paper published in Science, its autonomous work is like that of a human’s, but much faster.
“With Biomni, you can now execute complex tasks in minutes that normally take you weeks,” Biomni states when you load it up. “Design proteins. Analyze multi-omics data. Read and summarize thousands of papers.”
The model contains Biomni-E1 – its execution environment, armed with 68 specialized biomedical databases, 108 software packages and 82 tools, according to the website on launch – and Biomni-A1, the agent, or “co-scientist”.
“Biomedical research is increasingly constrained by repetitive, fragmented workflows that slow discovery,” the team writes. “We introduce Biomni, a general-purpose biomedical artificial intelligence agent that autonomously executes diverse research tasks. To map the biomedical action space, Biomni’s action-discovery agent mines tools, databases, and protocols from thousands of publications across 25 domains, building a unified agentic environment.
“Systematic benchmarking shows strong generalization across heterogeneous tasks – causal gene prioritization, drug repurposing, rare-disease diagnosis, microbiome analysis, and molecular cloning – without task-specific tuning,” they add. “Biomni envisions artificial intelligence augmenting human scientists and accelerating discovery.”
In one real-world example, a user uploaded more than 450 files of continuous blood-sugar monitoring, food intake and physical activity data and asked Biomni to analyze it and come up with “interesting and plausible” hypotheses. Some 40 minutes later, the AI had sorted the data, cleaned it up, created visual assets for the work and delivered on those observations based on patterns it picked up.
Leskovec says this job, which could be handled while the scientist was on their lunch break, would have taken a human 60 or more hours to complete.
While we may write about scientific discoveries daily, behind the scenes, research is incredibly slow to advance. And expensive to fund. However, it’s not due to a lack of talent in the field; instead, it’s hours upon hours of digesting literature, gathering and assessing data, writing code, investigating various avenues, identifying patterns, and much more.
Leskovec points out an interesting conundrum, too: as more scientific knowledge is produced, the slower advancements become.
“The hurdle in biomedical science is not intelligence or ideas; it is mechanics,” Leskovec adds. “It’s this laborious stuff that slows innovation. Biomni can do this work in minutes.”
The researchers do note that Biomni has some limitations, however. They note that its current iteration only covers part of biomedical research, with many key areas yet to be tested. And in complex multi-step tasks, something AI has not always performed well on when given total autonomy, Biomni still needs clear, structured prompts from a human scientist.
And they admit that the model has its strengths and weaknesses in its tasks. While it’s excellent at cloning molecules and querying databases, it’s not there yet when it comes to careful clinical judgment or experimental reasoning.
As such, Biomni is not designed to replace humans in the lab, but free them up to work on the less laborious tasks involved in their research.
“This is not about machines taking over science, but more about machines becoming a powerful new partner to augment human researchers,” says Huang. “With Biomni, scientists have a fast and tireless collaborator that empowers them to focus on the important work of science.”
Stanford notes that a Biomni prototype is already being used in more than 10,000 labs, making it the most widely utilized AI co-scientist system in the biomedical field.
“Biomni is my first research project that has gained wide use by real biologists,” Huang adds. “To have that impact on how biologists are doing their work has been rewarding. I look forward to seeing where Biomni goes from here.”
The research was published in the journal Science.
Source: Biomni and Stanford University
