AI-generated phages that work: ChatGPT for phage biologists?

Can AI design living organisms from scratch? Samuel King and Claudia Driscoll from the Arc Institute used genome language models to generate functional phages—and 16 of them actually worked in the lab. Not only that, they killed bacteria equally or better than the natural phage the team used as a template.

In this episode, we dig into how they did it, what it means for phage research, and why this could be a new way to explore evolution and design genomes.

What we covered:

• How genome language models work (ChatGPT, but for DNA)
• Training on millions of phage genomes with Evo-1 and Evo-2
• The creativity (from biologists!) required to figure out how best to filter generated sequences
• Going from 4000 selected sequences in silico, to 300+ synthesized candidates, to 16 working phages
• Fresh phage lab protocols for new ways to look at phage fitness
• Phage "personalities" that emerged from the generated candidates
• Watching recombination occur among a cocktail of designed phages
• Cost realities: hundreds of dollars per 5kb genome candidate, but emerging ways to reduce it exist!
• Good news: All tools are open source and free to use!

The preprint:
Samuel H. King, Claudia L. Driscoll, David B. Li, Daniel Guo, Aditi T. Merchant, Garyk Brixi, Max E. Wilkinson, Brian L. Hie (2025-09-17). Generative design of novel bacteriophages with genome language models | bioRxiv. biorxiv.org. Retrieved November 8, 2025, from https://www.biorxiv.org/content/10.11...

More resources:
• Arc Institute Evo browser interface: https://arcinstitute.org/tools/evo
• GitHub (open source code): Evo2: https://github.com/ArcInstitute/evo2)
• Hugging Face (model downloads): https://huggingface.co/arcinstitute

Guests:
Samuel King: PhD candidate, Stanford/Arc Institute, Brian Hie's lab; follow @samuelhking on X
Claudia Driscoll: Postdoc, Arc Institute, Brian Hie's lab, follow @driscoll_cl on X

Also follow @BrianHie, @arcinstitute, @stanford on X for more from this team!