e typically think of AI models as built around language. But one startup is creating models built around the "grammar of our DNA."
Radical Numerics, an AI research lab spun out from Stanford University on the quest to build "general biological intelligence," emerged from stealth last week with $50 million in seed funding. Additionally, the company announced its latest model: Omnii, a "genome language model" built for researchers and biologists, in research preview.
"Most of the DNA in the human genome, we still don't actually understand what it does. We don't understand the grammar, so to speak, of our DNA," Eric Nguyen, CEO and co-founder of Radical Numerics, told The Deep View. "We thought this was a way to accelerate our understanding of disease, and hopefully to make treatments to cure disease."
Though the company has only just revealed itself to the public, it's been at work for almost a year, said Nguyen, and has already created and released models.
- Evo and Evo 2 , its first two models, are the first AI models capable of both reading and writing DNA at scale. The project is entirely open source, Nguyen said, and was used to design novel CRISPR systems and create the first complete AI-designed genome, called a bacteriophage. - Because the field is so nascent, the company released these models open source as a means of pushing innovation and getting it in the hands of researchers "in an open way," he said. "We felt that we had to show people the recipe and make it open source to actually drive adoption."
However, amid a growing discussion of AI's potential use in developing bioweapons, Nguyen said that the company approaches its technology in a "dual prong" way, considering the tech for both biological design and biological defense. It's why the company decided to keep Omnii proprietary, compared to its Evo predecessors: Being able to control who has access to its models is vital. "We don't want this to be an academic toy anymore," Nguyen said.
"We realized that this technology has this tremendous potential, obviously, to understand human health, but it can be potentially misused to create something very dangerous," he added.
Our Deeper View #
Radical Numerics joins a growing movement targeting bioscience as a use case for AI. OpenAI introduced GPT-Rosalind specifically for life sciences research and drug discovery, and Anthropic has its own version of Claude for Life Sciences. Additionally, each of these initiatives has a precedent: Alphafold, Google DeepMind's AI project predicting the 3D structures of proteins, which won the Nobel Prize in 2024. And there's a reason this particular niche is so attractive — One of the big pieces of the utopian vision for AI is to provide researchers with the tools to exponentially accelerate scientific discovery, positioning the tech as a panacea for all of the hurdles that ail researchers. Still, Radical Numerics acknowledges the biggest risk: That a tool of such power can be used to tip the scales in a dangerous direction, if not controlled.