AI Designed the First Functional Viral Genomes From Scratch — 16 New Bacteriophages That Beat Nature

Evo 2 is a genomic foundation model trained on DNA sequences across bacteria, archaea, plants, animals, and humans, operating at single-nucleotide resolution . The researchers fine-tuned both Evo 1 and Evo 2 on bacteriophage genomes to produce whole-genome sequences with realistic genetic architecture and desirable host targeting .

The pipeline was massive in scale. The models generated approximately 700,000 candidate viral genome sequences . Computational filters winnowed those down to 302 promising candidates . Of those, 285 were chemically synthesized — printed as physical DNA — and introduced into E. coli cells .

16 of the 285 synthesized genomes produced functional, replicating phages — a success rate of about 5.6% . The study establishes the first measurable "productive yield" benchmark for genome-scale generative biology .

Bacteriophage Results: Outperforming Nature

All 16 viable phages infected and lysed (killed) E. coli in lab dishes, demonstrating full replication, infection, and host cell lysis . Several of the AI-designed phages outperformed wild-type ΦX174 — the natural template virus — in head-to-head comparisons of growth rate and lysis speed . Some designs contained over 50 amino acid changes compared to the natural virus, a level of divergence far exceeding what random mutation would typically produce .

One of the phages encoded a DNA packaging protein not found in any known organism on Earth — a completely novel biological structure invented by the AI .

A cocktail of the 16 phages overcame resistance across multiple E. coli strains, including virus-resistant variants that had evolved immunity to individual phages .

Implications for Antibiotic-Resistant Infection Treatments

The ability to rapidly generate novel phages on demand could bypass the evolutionary arms race in which bacteria develop resistance to individual phage strains . Because phage cocktails can target bacteria through multiple mechanisms simultaneously, the AI-generated phages offer a promising weapon against multidrug-resistant infections .

The study establishes that genome language models can experimentally produce functional viruses, opening the door to rapidly prototyping new therapeutic phages for specific bacterial pathogens . Brian Hie noted the program's ability to incorporate evolutionary diversity could make phage therapy a more effective alternative to traditional antibiotics .

Biosecurity Concerns: A Governance Gap

In an accompanying commentary in Science, Dr. Thomas Inglesby and Dr. Moritz Hanke of the Johns Hopkins Center for Health Security wrote: "Although this is promising for life sciences applications, it also raises urgent biosafety and biosecurity questions. The ability to compose viral genomes using generative AI now exists; the governance to ensure these capabilities are used safely and securely does not" .

The primary safeguard — the International Gene Synthesis Consortium's screening system that checks DNA orders against known pathogen sequences — is voluntary, not legally required, and was never designed to catch sequences that have no match in any natural database .

The Arc Institute took safety precautions by deliberately excluding eukaryotic viral pathogens (viruses that infect humans or other animals) from Evo 2's training data . However, experts warned that the same AI methods could be applied by bad actors to generate sequences for dangerous pathogens, and that AI-designed genomes could be created too quickly for current surveillance systems to detect or stop them .

A Nature commentary also highlighted that while these tools are developed for beneficial drug discovery, they could lower the barrier to creating novel biological threats .