Blab AI
Stanford AI Advances Accelerate Biological and Astronomical Research
Stanford University’s Human‑Centered AI (HAI) institute has unveiled a quartet of artificial‑intelligence tools that promise to reshape the way scientists explore biology and astronomy. The new systems—Evo 2, Biomni, a virtual‑cell foundation model, and the Center for Decoding the Universe—show how large language models and autonomous agents can speed discovery, automate routine tasks, and generate fresh hypotheses.
Evo 2 is a DNA‑language model that entered the public domain in February 2025. Trained on 9.3 trillion base pairs drawn from a curated atlas that includes genomes from every domain of life—including humans—Evo 2 pairs a 40‑billion‑parameter neural network with a context window capable of handling up to one million nucleotides at once. The open‑source model is reachable via a web interface that accepts a short DNA sequence and returns an autocompleted gene. The output can match a natural gene exactly or propose a mutated variant for researchers to assess in medicine or agriculture. The project, funded by a Hoffman‑Yee Research Grant, was led by Stanford assistant professor Brian Hie.
Biomni is a biomedical AI agent that stitches together hundreds of specialized tools, databases, and software packages into a single, task‑agnostic research environment. It can write Python code, pull scientific literature, and run complex analyses without the need for task‑specific tuning. According to a report, 15,000 scientists have already employed Biomni to automate 100,000 scientific workflows, ranging from gene prioritization to drug repurposing. The team includes researchers from Stanford, Genentech, Arc Institute, the University of Washington, Princeton, and UCSF.
Separately, Stanford is building a human‑centered foundation model that can simulate individual cells. Associate professor Emma Lundberg describes the goal as creating a digital twin of a person’s cells that could be used to test drug responses before clinical trials. The project must overcome two major hurdles: first, integrating diverse biological data types—DNA, RNA, protein sequences, protein structures, cellular images, and literature; second, providing a chat interface that lets biologists upload their own datasets. The team has prototyped a chat system called Biomni to meet this need.
To gauge how well large language models can generate research ideas, Stanford scholars ran a head‑to‑head study with 100 human experts. The experts reviewed 49 ideas across seven topics, including bias, coding, and safety. The AI agent produced 4,000 seed ideas per topic. Reviewers scored each idea on novelty, excitement, feasibility, and expected effectiveness. The study found that AI‑generated ideas were judged more novel than human ideas, but many lacked feasibility. Lead author Chenglei Si noted that while the language model shows strong technical creativity, humans still produce more practical proposals.
In astronomy, the Rubin Observatory in Chile is launching a ten‑year sky survey that will generate millions of images and real‑time alerts for thousands of scientists worldwide. Risa Wechsler, director of Stanford’s Kavli Institute for Particle Astrophysics and Cosmology (KIPAC), says the survey will provide unprecedented data for studying dark energy and stellar evolution. The Center for Decoding the Universe, a joint initiative between KIPAC and Stanford Data Science, is developing AI agents and frontier models to extract insights from the multimodal datasets. Faculty at the center are exploring how generative models can simulate complex astrophysical systems.
Together, these projects illustrate a broader trend in which AI becomes a collaborator in scientific research. By automating routine tasks, generating hypotheses, and providing virtual testbeds, AI tools can accelerate discovery across disciplines. The Stanford initiatives also highlight the importance of guardrails—ensuring data quality and addressing bias—to maintain scientific rigor.
Today, AI is already embedded in real research workflows. Evo 2 is being used by geneticists to design novel gene sequences; Biomni is streamlining drug‑discovery pipelines; the virtual‑cell model remains in development but aims to enable personalized medicine; and the Center for Decoding the Universe is preparing to handle the Rubin Observatory’s data deluge. As these tools mature, they are likely to influence funding priorities, research collaborations, and the pace of scientific breakthroughs.
The next few years will see continued investment in AI for science. Stanford’s HAI and Data Science initiatives are expanding their reach, and other institutions are expected to adopt similar models. However, the field must also address challenges related to reproducibility, data ownership, and equitable access to ensure that the benefits of AI‑assisted research are broadly shared.
Evo 2 is a DNA‑language model that entered the public domain in February 2025. Trained on 9.3 trillion base pairs drawn from a curated atlas that includes genomes from every domain of life—including humans—Evo 2 pairs a 40‑billion‑parameter neural network with a context window capable of handling up to one million nucleotides at once. The open‑source model is reachable via a web interface that accepts a short DNA sequence and returns an autocompleted gene. The output can match a natural gene exactly or propose a mutated variant for researchers to assess in medicine or agriculture. The project, funded by a Hoffman‑Yee Research Grant, was led by Stanford assistant professor Brian Hie.
Biomni is a biomedical AI agent that stitches together hundreds of specialized tools, databases, and software packages into a single, task‑agnostic research environment. It can write Python code, pull scientific literature, and run complex analyses without the need for task‑specific tuning. According to a report, 15,000 scientists have already employed Biomni to automate 100,000 scientific workflows, ranging from gene prioritization to drug repurposing. The team includes researchers from Stanford, Genentech, Arc Institute, the University of Washington, Princeton, and UCSF.
Separately, Stanford is building a human‑centered foundation model that can simulate individual cells. Associate professor Emma Lundberg describes the goal as creating a digital twin of a person’s cells that could be used to test drug responses before clinical trials. The project must overcome two major hurdles: first, integrating diverse biological data types—DNA, RNA, protein sequences, protein structures, cellular images, and literature; second, providing a chat interface that lets biologists upload their own datasets. The team has prototyped a chat system called Biomni to meet this need.
To gauge how well large language models can generate research ideas, Stanford scholars ran a head‑to‑head study with 100 human experts. The experts reviewed 49 ideas across seven topics, including bias, coding, and safety. The AI agent produced 4,000 seed ideas per topic. Reviewers scored each idea on novelty, excitement, feasibility, and expected effectiveness. The study found that AI‑generated ideas were judged more novel than human ideas, but many lacked feasibility. Lead author Chenglei Si noted that while the language model shows strong technical creativity, humans still produce more practical proposals.
In astronomy, the Rubin Observatory in Chile is launching a ten‑year sky survey that will generate millions of images and real‑time alerts for thousands of scientists worldwide. Risa Wechsler, director of Stanford’s Kavli Institute for Particle Astrophysics and Cosmology (KIPAC), says the survey will provide unprecedented data for studying dark energy and stellar evolution. The Center for Decoding the Universe, a joint initiative between KIPAC and Stanford Data Science, is developing AI agents and frontier models to extract insights from the multimodal datasets. Faculty at the center are exploring how generative models can simulate complex astrophysical systems.
Together, these projects illustrate a broader trend in which AI becomes a collaborator in scientific research. By automating routine tasks, generating hypotheses, and providing virtual testbeds, AI tools can accelerate discovery across disciplines. The Stanford initiatives also highlight the importance of guardrails—ensuring data quality and addressing bias—to maintain scientific rigor.
Today, AI is already embedded in real research workflows. Evo 2 is being used by geneticists to design novel gene sequences; Biomni is streamlining drug‑discovery pipelines; the virtual‑cell model remains in development but aims to enable personalized medicine; and the Center for Decoding the Universe is preparing to handle the Rubin Observatory’s data deluge. As these tools mature, they are likely to influence funding priorities, research collaborations, and the pace of scientific breakthroughs.
The next few years will see continued investment in AI for science. Stanford’s HAI and Data Science initiatives are expanding their reach, and other institutions are expected to adopt similar models. However, the field must also address challenges related to reproducibility, data ownership, and equitable access to ensure that the benefits of AI‑assisted research are broadly shared.
