Researchers at the Arc Institute have created the first functional virus genomes entirely designed by artificial intelligence. Of 285 AI-generated designs, 16 produced viable bacteriophages that successfully infected and killed E. coli bacteria. When combined into cocktails, these phages overcame antibiotic resistance where natural viruses failed.
What the Researchers Actually Did
Samuel King and Brian Hie used Evo, a genomic foundation model trained on over 2 million bacteriophage genomes, to generate entirely new genetic sequences. Their target was ΦX174, a well-characterized phage with 5,386 nucleotides and 11 overlapping genes. This choice was deliberate: ΦX174’s complexity creates stringent constraints that make random generation essentially impossible.
The process involved fine-tuning Evo on 14,466 Microviridae sequences, then generating candidate genomes with carefully tuned sampling parameters. The team synthesized these designs using Gibson assembly, transformed them into E. coli, and monitored for signs of infection in 96-well plates.
Results came quickly. Within 2-3 hours, successful phages caused rapid drops in bacterial density. Each functional genome contained between 67 and 392 mutations compared to the nearest known natural sequence. One phage, dubbed Evo-Φ2147, was so different from anything in databases (below 95% sequence identity) that it would technically qualify as a new species.
The Antibiotic Resistance Test
The most significant finding: AI-designed phage cocktails overcame bacterial resistance that stopped natural phages cold.
When researchers created three E. coli strains resistant to wild-type ΦX174, the original phage failed completely to overcome any of them. But cocktails of AI-designed phages cleared resistance in all three strains within 1-5 passages. The successful phages were “mosaic genomes derived from multiple AI designs through recombination,” meaning the bacteria couldn’t evolve fast enough to escape.
This matters because antibiotic resistance kills an estimated 1.27 million people annually, and phage therapy has struggled to gain traction partly because natural phages are too limited in their capabilities.
What This Means
British molecular biologist Adrian Woolfson called the achievement “a massive, consequential moment” marking a shift “from a Darwinian world into a post-Darwinian landscape” where life can be authored rather than simply inherited.
That’s not hyperbole. Previous attempts at phage engineering worked by modifying existing viruses. This approach generates functional genomes from scratch, including novel solutions like one phage (Evo-Φ36) that successfully incorporated a DNA-packaging protein from a distantly related virus - something human engineers had failed to achieve.
The implications for medicine are substantial. If AI can generate diverse phage libraries on demand, personalized cocktails could be designed to target specific infections, potentially reviving phage therapy as a serious alternative to failing antibiotics.
The Fine Print
The 5.6% success rate (16 of 285 designs) shows AI genome design is far from reliable. Most generated sequences were non-functional, and the researchers note that “complexity of genome-scale design introduces new constraints and failure modes absent in single-protein design.”
Biosecurity concerns are being taken seriously. The team conducted experiments in dedicated biosafety cabinets using only non-pathogenic laboratory E. coli strains. Critically, Evo was deliberately trained without human viral sequences, and the template-based method preserves host specificity through conserved spike proteins - meaning these phages can’t infect humans.
Still, the technology exists now. As genome synthesis costs continue to fall, the barrier to creating novel pathogens drops with them. The researchers are calling for regulation before this moves beyond the lab, a view echoed by genome pioneer George Church, who warned that “extreme caution” is warranted.
This is a genuine breakthrough with genuine risks. The question is whether regulatory frameworks can keep pace with the technology.