AI can now design functional viruses – not the computer kind, either

Inject this synthetic phage into E. coli and it kills better than the real thing

A group of Stanford bioengineers claim that they've created synthetic bacteriophages using AI-generated designs that not only work in the real world, but are far more infectious than their naturally-occurring counterparts. 

The team, led by Stanford chemical engineering professor Brian Hie, posted a paper to the preprint service bioRxiv Wednesday that details their use of the Evo 1 and Evo 2 models from the Arc Institute (Hie contributed to the design of both Evo models). According to their findings, and Hie's comments to The Register, this is the first time an AI-generated genome has been produced in the real world and found functional in tests. 

Bacteriophages, which are viruses that infect bacteria cells, have been widely studied since their discovery in the early 20th century. Of those, one particular bacteriophage, dubbed ΦX174, has been extensively studied and genetically mapped. It's that particular bacteriophage that served as the starting point for Hie and his team's research. ΦX174 specifically targets E. coli bacteria.

Neither Evo model can simply produce a functional virus genome without a lot of help. In order to develop ΦX174 variations that would have beneficial mutations, Hie and his team introduced both Evo models to additional genetic samples, specifically engineered prompts "with ΦX174-specific sequences," and used inference-time guidance to tweak outputs. 

The end result of the AI portion of the project was a pool of 302 candidate genomes, 285 of which were able to generate full genomes. Of those, 16 were found to inhibit the growth of E. coli bacteria. 

Those synthetic bacteriophages didn't just inhibit it as well as ΦX174, though: They blew its infectious capabilities out of the water. Across three experiments, ΦX174 only managed to rank as high as the third most effective E. coli killer, and didn't even place in the top five most infectious bacteriophages in one of the three tests.

The infectiousness of ΦX174 barely even holds a candle to the top performer, Evo-Φ69, which showed an expansion rate between 16 fold and 65 fold over a six-hour infection period. Poor ΦX174 only showed increases between 1.3 fold and fourfold over the same period. 

That said, the bacteriophage genomes that Hie and his team created aren't able to do much without a bit of a push from human researchers since they're not fully-formed bacteriophages themselves.

Viruses replicate by injecting their genomes into target cells, which are hijacked to create the rest of the parts needed for a new virus to spread to additional cells. It's those cell-hijacking genomes the team created, and they had to be injected into E. coli cells before they could wreak havoc. Once that initial work is performed, however, Hie told us it's a simple matter to make more. 

"Once the phage gets made once, we can just save the phage itself, which can be grown up in the lab for phage therapy," Hie told us in an email. 

Such therapies, Hie told us, could be used to target multidrug resistant bacterial infections – a fact that was confirmed during the research project. Hie and his team evolved three strains of E. coli resistant to ΦX174 for their project, and found that mutations in the artificial bacteriophage were able to overcome bacterial resistance through recombination of multiple synthetic phage strains. 

That, the team said, suggests that generative AI could be used to "produce genetically diverse phage cocktails that could translate into improved therapeutic efficacy." In other words, AI can easily overcome the limitations of natural biology.

"We envision that techniques developed in this work offer a path toward designing more complex biological systems with desirable functions, potentially including the larger genomes of living organisms," the team said in their paper. Hie told us that his team is now working toward finding clinical applications for both human bacterial infections and those that target crops. 

The paper is now under peer review, Hie told us.

Does this sound like a Crichton novel to you?

The idea of researchers creating synthetic virus genomes that are far more infectious than the real thing sounds like the plot of an apocalyptic novel that ends with humanity dying from its own hubris. Biotechnology journalist Niko McCarty at the Asimov Press rasied that concern in a story about Hie's team's discovery. 

According to the rules of phage taxonomy, McCarty said, several of the AI-generated phages developed in this project qualify as entirely new species. What's to stop a well-motivated and funded person, or more likely a nation-state, from taking this work and the open-source models that helped it succeed, and turn that toward creating deadly human viruses? 

"This paper will probably be alarming to folks in the biosecurity community," McCarty wrote. "The HIV genome is only about 10,000 bases in length (not much larger than bacteriophages) and the coronavirus genome is about 30,000 bases." 

It's worth considering that the Evo models didn't include human virus genome data in their pretraining data. As we noted earlier, Hie's team had to do extensive work – including prompt engineering and the addition of lots more genome training data – to succeed. The same, theoretically, could be done by a bad actor targeting humans. 

Hie's not concerned. 

"We did have to do a lot of work purely on the computational end to get coherent generation from the model, and that was with Evo 1 and 2 being trained directly on phages," Hie explained to us. "Since Evo 1 and 2 were not trained on human viruses, we expect this to be a lot harder, and this is before doing any of the experimental work." 

Hie explained that a common litmus test in the biosecurity community is whether something lowers the barrier of access to designing a bioweapon, which he argues this does not do because of all the legwork required by his team to get from Evo to artificial bacteriophages. 

"If you wanted to design a bioweapon with AI, it would just be much harder than taking something from nature," Hie said. 

So it's not that it can't be done – it'd just be impractical. For now, at least. Let's hope getting to that next stage takes a while. ®