AI designs functional viruses for first time, raising medical and

AI generates first functional viral genomes

Stanford University and Arc Institute researchers used a generative AI model called Evo to design complete viral genomes from scratch. Of 302 synthesized designs, 16 produced functional bacteriophages that infected and replicated within E. coli bacteria, with some outperforming natural counterparts. The work, published in Science, marks the first successful AI-designed organisms capable of cellular replication.

How the viral design system works

The Evo model operates similarly to large language models but predicts DNA sequences instead of words. Trained on nine trillion nucleotides from millions of organisms, it learned to generate viable viral genomes after fine-tuning on 15,000 relatives of the Phi X-174 bacteriophage. Researchers selected candidates based on genetic plausibility before lab synthesis and testing.

Medical potential for antibiotic resistance

All designed viruses targeted E. coli, demonstrating potential for phage therapy - using viruses to kill specific bacteria. Study co-author Samuel King noted this approach could address antibiotic-resistant infections, as phages can be engineered against pathogens where drugs fail. The AI reportedly created some phages that overcame bacterial resistance mechanisms.

Built-in safety measures

Researchers implemented multiple safeguards: training data excluded human-infecting viruses, experiments used non-pathogenic E. coli, and work occurred in a secure lab. The team chose bacteriophages specifically because they cannot infect human cells. Lead researcher Brian Hie emphasized these were deliberate constraints to mitigate risks.

Emerging biosecurity concerns

Independent experts warn the technology could be repurposed. In a Science commentary, Johns Hopkins biosecurity researchers noted AI now enables viral genome design without the expertise traditionally required. They highlighted the lack of governance frameworks for preventing misuse with human pathogens, calling this an "urgent" gap.

Open-source access complicates control

Unlike corporate AI projects, Evo's code is publicly available. While the Stanford team trained it only on safe viral data, others could retrain models with different datasets. This accessibility raises questions about preventing malicious use while preserving research benefits.

Next steps for the technology

The researchers suggest focusing initial applications on well-characterized systems like bacteriophages. For human medical use, they propose strict oversight protocols and physical containment. Independent analysts recommend international standards for AI-generated organisms, similar to gain-of-function research rules.

Readers working in biotech or AI policy can review the Science publication and accompanying commentary to evaluate both the technical methods and proposed governance approaches.