Arc Institute builds first viable AI-generated viral genomes

Designing a single protein with AI is already hard, but generating a complete, self-replicating biological organism from scratch has always been biology's moonshot.

Arc Institute researchers just pulled it off. Using their genomic foundation model Evo, the team designed and physically synthesized complete bacteriophage genomes that are fully functional, successfully infecting and lysing target bacteria. It marks the first time an AI-designed biological genome has come to life in the lab.

From reading to writing to AI-generating life

The team chose bacteriophage ΦX174 as their canvas, a virus made of 5,386 nucleotides encoding 11 genes. It has deep historical roots: in 1977, Fred Sanger chose it to produce the first-ever sequenced genome; in 2003, Craig Venter chemically synthesized it from scratch to prove whole genomes could be written. In 2025, Arc made it the first template for generative AI.

ΦX174 is notoriously complex because its genes overlap, meaning multiple proteins share the exact same sequence of DNA letters. Conventional gene predictors missed roughly a third of its genes. The researchers first built a custom annotation tool to reliably identify all 11 genes, then fine-tuned Evo on 14,466 Microviridae sequences. By prompting the model carefully, they coaxed it into generating genomes that conserved critical machinery while exploring uncharted evolutionary paths.

Out of 285 AI-generated genomes synthesized in the lab, 16 successfully formed viable phages that rapidly destroyed host bacteria. These working viruses carried between 67 and 392 mutations compared to their closest natural relatives, with 13 genomes possessing mutations never seen in nature. One candidate, Evo-Φ2147, had 392 mutations and only 93% nucleotide identity to wild phages, technically classifying it as an entirely new synthetic species.

Another design, Evo-Φ36, seamlessly swapped in a structural protein from a distant cousin, phage G4. Previous rational human engineering failed to make that swap work, but cryo-EM imaging revealed Evo automatically made subtle compensatory mutations across the capsid to stabilize the new protein.

When tested against bacteria engineered to resist standard phages, the AI-designed variants recombined and overcame the resistance within just 1 to 5 passages, whereas the natural wild-type phage failed completely.

Why it matters

This is a massive conceptual leap from structural biology to living systems. Until now, AI in biology was mostly confined to isolated molecules like antibodies or single enzymes. Evo proves that large language models can grasp the nuanced, overlapping feedback loops required to run an entire organism, turning biology from a process of discovery into an engineerable computational stack.

For medicine, the timing could not be better. With antibiotic-resistant superbugs threatening global healthcare, finding natural phages to treat stubborn infections is currently an agonizing trial-and-error hunt. We think generating diverse synthetic phage cocktails on demand fundamentally tilts the evolutionary race back in our favor, allowing researchers to deploy custom viral therapies before bacteria can adapt.

At the same time, this milestone will inevitably set off bio-risk alarms. The Arc team took strict precautions, filtering human viruses from Evo's training corpus and restricting the phages to harmless lab strains of E. coli. But once the playbook for generating functional viral genomes is out in the wild, ensuring that bad actors cannot redirect similar models toward harmful pathogens will be one of the hardest regulatory challenges the industry has ever faced.