A study from Stanford University and the Arc Institute, published in Nature, demonstrates that artificial intelligence (AI) can design complete bacteriophage genomes rather than only predicting protein structures. The researchers synthesized 302 completely AI-written bacteriophage designs; of these, 16 yielded functional viruses that infected, replicated within, and killed Escherichia coli.
What the team did
- The work was produced by researchers at Stanford University and the Arc Institute and reported in Nature.
- An AI system referred to as “Evo” generated full genomic sequences for bacteriophages.
- The team synthesized 302 of these AI-designed genomes in the laboratory.
- Sixteen of the synthesized designs resulted in particles that behaved as functioning viruses: they could infect E. coli, replicate, and lyse bacterial cells.
Notable finding: a capsid protein with no known natural relative
Among the AI-designed genomes was at least one phage carrying a capsid protein that did not match any known natural homolog. The authors interpret this as evidence that the model is not merely recombining existing natural sequences, but can explore regions of sequence and structural design space that nature has not sampled.
Why this matters
- Technological step change: while prior high-profile AI contributions in biology — for example Google’s AlphaFold — focused on predicting protein folding, this work shows AI moving from prediction to generative design at the genomic level and to products that function in wet labs.
- Acceleration and risk: AI-enabled design of functioning biological agents could speed up areas such as drug discovery and synthetic biology. At the same time it raises biosecurity and ethical concerns, because the production and dissemination of designed organisms require oversight and safeguards.
Summary
According to the Nature paper by Stanford University and the Arc Institute teams, AI can write full bacteriophage genomes that produce viable viruses in laboratory conditions. Out of 302 synthesized AI-designed genomes, 16 produced functional viruses that infected, replicated in, and killed Escherichia coli, and one design included a capsid protein without any known natural relative, indicating AI exploration beyond natural sequence space.
