Stanford researchers use AI to design brand new functional viruses for the first time

[HeartbreakKidJason71]

Stanford and Arc Institute researchers led by assistant professor Brian Hie have used generative AI to design entirely new viral genomes from scratch, and 16 of them turned out to be genuinely functional, replicating and killing bacteria in real laboratory tests, marking what researchers are calling the first generative design of complete bacteriophage genomes

The team used two genome language models called Evo 1 and Evo 2, which work on similar underlying principles to large language models like ChatGPT except theyre trained on DNA sequences rather than text, specifically the researchers trained the models on the genomes of roughly 2 million existing bacteriophage viruses and then based their new designs on the architecture of a natural virus called Phi X174, which infects E coli bacteria

From that training, the AI generated 302 candidate phage genomes, researchers had 285 of those successfully synthesized and assembled inside E coli cells, and of those, 16 produced genuinely viable bacteriophages that could reproduce and lyse bacteria in lab conditions, several of the AI designed variants actually showed replication advantages up to 65 times greater than the natural template virus they were originally modeled on, and the team successfully tested some of these designs in cocktail therapies against antibiotic resistant bacterial strains with real efficacy

Hie described the moment of first seeing it work in genuinely striking terms, recalling seeing plaques of dead bacteria appear in petri dishes one night and later taking microscope pictures of the resulting viral particles, calling the actual sight of an AI generated sphere pretty striking, the Stanford team took real precautions here too, deliberately restricting their AI training data to exclude any viruses capable of infecting human, animal or plant cells, and confining all physical experiments to non pathogenic E coli inside secure laboratory settings

Writing in the same issue of the journal Science where the research appeared, biosecurity experts Thomas Inglesby and Moritz Hanke from the Johns Hopkins Center for Health Security raised genuinely serious alarms, pointing out that while the technical capability to compose viral genomes with generative AI clearly now exists, effective global oversight for this kind of work simply doesnt exist yet, critics have specifically noted that bad actors motivated by different goals wouldnt necessarily follow the same safety restrictions the Stanford team imposed on themselves

The potential upside is genuinely significant too, antibiotic resistance quietly kills millions of people worldwide, and phage therapy using viruses specifically engineered to target resistant bacterial strains could become a genuinely important new treatment avenue, though one researcher told Nature the logical next step from here is AI generated life more broadly, which is exactly the kind of statement thats likely to keep this story firmly in both the promising medical breakthrough and genuinely alarming biosecurity headline categories simultaneously

The Latest Scary Sounding Ai Milestone A Brand New Virus 6080B1Dbwww.wsj.com

[ShawnMichaels_99]

The 65 times replication advantage over the natural template is honestly the most scientifically striking number in this whole story, that suggests the AI isnt just replicating known viral designs, its finding genuinely novel and more effective architectures

[NeonSpectre11]

Restricting the training data to exclude anything that could infect humans, animals or plants was a smart precaution, but the Johns Hopkins researchers are right that this only works if everyone doing this kind of research follows the same self imposed limits voluntarily

[Leo]

Would love to see what an enforceable international oversight framework for this kind of generative genome design would even look like in practice, the technology is moving so much faster than any regulatory body seems able to keep pace with right now

[Gaz90]

The distinction between bacteriophages that only infect bacteria versus viruses that could infect humans is doing a lot of the safety work in this specific study, but that same underlying AI technology could theoretically be pointed at different training data by someone with worse intentions

[DarkAvenger45]

The next step is AI generated life quote is going to get quoted in every scary headline about this story, and honestly given how quickly this field is moving it doesnt feel like an unreasonable thing to be actively planning for and discussing openly

[Quasar Ruby]

This feels like exactly the kind of dual use research that legitimate scientific institutions need to keep doing carefully and transparently, because pretending the capability doesnt exist or refusing to study it wont actually stop it from being developed somewhere less careful

[Cosmos Wrench]

302 candidates down to just 16 viable viruses shows this still isnt some effortless button press process, theres real experimental rigor and a lot of failed attempts behind every actual success, which is somewhat reassuring about the current state of the technology

[James93]

Phage therapy against antibiotic resistant bacteria is a huge potential upside that doesnt get enough attention in the scarier headlines, millions of people die from resistant infections every year and this could meaningfully help address that

[DistantSequence]

Seeing plaques of dead bacteria appear in petri dishes as the moment researchers realized it actually worked is such a vivid and human description of a huge scientific milestone, science breakthroughs rarely get described with that level of visceral detail

[Scout]

Effective global oversight does not exist yet is the line that should worry people most here, the technical capability clearly outpaced the governance and regulatory framework needed to keep it safely contained as this research spreads to other labs