AI that designed its own viruses

In a paper published in Science on 6 August, a team at Stanford led by computational biologist Brian Hie used two generative AI models, Evo 1 and Evo 2, to design complete viral genomes from scratch. The viruses in question are bacteriophages, which attack bacteria rather than people, and the model, trained on trillions of nucleotides of genetic sequence, works the way a large language model does, except that instead of predicting the next word it predicts the next base pair. Given a starting point, it proposes genomes; the team selected about 700,000 candidate designs, built 285 of them in the lab and found that sixteen of them successfully infected and killed E. coli in the dish. Some of the synthetic phages appeared to multiply and spread better than the natural phage they were modelled on.

The starting point was Phi X-174, a well-studied phage with a genome of roughly 5,400 base pairs, about a million times smaller than a human’s. The sixteen working designs were close enough to their model that a cocktail of all sixteen defeated two strains of antibiotic-resistant E. coli, something neither the natural phage nor a cocktail of natural phages managed. That is the practical hook: antibiotic resistance kills an estimated 4.7 million people a year, and phages are one of the few weapons left that bacteria have not fully learned to dodge. A single phage fails when the bacteria evolve around it; a mixture of genetically distinct phages is harder to out-evolve, which is the argument Hie makes for the cocktail approach.

The careful parts matter as much as the impressive ones. The training data for Evo deliberately excludes genomes known to infect humans, animals, plants and fungi, so the model has no learned route to a virus that attacks people; the team also ran the experiments under containment conditions beyond what is standard for phage work, and published those procedures as a framework others can follow. The authors frame the study as proof that genome design has become a generative problem like any other, with the same pattern of falling costs and rising capability.

That is precisely why the paper has drawn two kinds of reaction. The first is excitement: a targeted phage for a resistant infection, designed in days rather than months, is a useful medical tool. The second is unease about where the same capability points. The technology that designed a 5,400-base-pair genome will not stay at 5,400 base pairs. Tom Ellis, a synthetic genome engineer at Imperial College London, pointed out that the Covid-19 virus genome is about six times longer than Phi X-174, and the difficulty of designing a genome grows far faster than its length, by perhaps a factor of a hundred for a sixfold increase. The gap is real, but it is a gap that closes, and biosafety frameworks built for today’s models will be tested against tomorrow’s.

The debate resembles the one that followed the 2011 experiments that made an influenza virus more transmissible, and the synthesis of the first complete bacterial genomes a few years later. Each time, the pattern has been the same: a capability is demonstrated, the dual-use question is asked, safeguards are promised, and the field moves on faster than the governance does. The pattern has held for more than a decade, and the speed of generative models makes the next round harder to manage, because the barrier to entry is no longer a well-funded lab.

None of this is a reason to abandon the line of work. Antibiotic resistance is a quiet catastrophe that kills more people each year than most wars, and phages are the best answer biology has offered to it so far. The argument is for the two things to develop together: the models, and the rules for handling what they make. The Science paper includes a biosafety framework in its supplementary materials; frameworks that sit inside papers are a start, but they only matter if they survive contact with the next model release. Sixteen working viruses is a small number, until it is not.