In laboratories where the boundary between biology and computation has grown thin, scientists have done something quietly extraordinary: they asked an artificial intelligence to imagine viruses that nature never made, and the viruses worked. Published in Science, the research describes how AI-designed bacteriophages — living medicines conjured from learned patterns in DNA — defeated antibiotic-resistant bacteria that natural viruses could not overcome. It is an early signal that medicine may soon be able to design its own evolutionary answers to one of its oldest and most dangerous adversaries
AI designs novel viruses to fight antibiotic-resistant bacteria
Design new phages on demand, tailored to attack whatever bacteria needs treating
Mark
Why does it matter that these viruses were designed by AI rather than found in nature?
Mimi
Because nature is slow and limited. If you need a phage to fight a specific resistant strain, you might search for years and find nothing. With AI, you can generate thousands of candidates in hours and test them immediately. You're not waiting for evolution to hand you the answer.
Mark
But aren't you worried about releasing viruses that have never existed before into the world?
Mimi
That's the right instinct, but these phages are extraordinarily narrow. They infect only bacteria—not human cells, not plants, not fungi. And they're designed to attack specific strains. It's more like a precision tool than a loose cannon.
Mark
The bacteria became resistant to the natural virus pretty quickly. Won't they do the same to these AI-designed ones?
Mimi
They might, eventually. But the cocktail approach buys you time. A mutation that defeats one phage doesn't defeat all 16. The bacteria would need to evolve multiple independent defenses simultaneously, which is far harder. And if resistance does emerge, you can design new phages to counter it.
Mark
So this is an arms race that never ends?
Mimi
In a sense, yes. But it's an arms race where one side can design new weapons in weeks instead of waiting for nature to provide them. That's a fundamental shift in how we might approach bacterial infections.
Mark
What happens next? Is this ready for patients?
Mimi
Not yet. This is proof of concept in a lab. The next steps are testing in animal models, then human trials. But the pathway is clear now. If it works at scale, phage therapy could become routine within a decade.
Le Pouls
- Antibiotic resistance is quietly becoming one of the deadliest forces in modern medicine, with infections like MRSA and tuberculosis outpacing the drugs designed to stop them.
- Finding natural bacteriophages capable of targeting specific resistant bacteria is slow, unreliable, and increasingly inadequate — the biological library nature offers is finite.
- Researchers trained an AI on millions of viral DNA sequences, generating 700,000 novel phage genomes that have never existed in nature, then chemically synthesized the most promising 285.
- Sixteen of those synthetic viruses came to life inside bacterial cells and functioned — not as copies of known viruses, but as genuinely new biological entities with novel gene combinations.
- When mixed into a cocktail and deployed against bacteria already resistant to natural phages, the AI-designed viruses overwhelmed the bacteria by attacking from multiple genetic angles simultaneously.
- The research points toward a future where custom viral medicines could be designed on demand in weeks rather than years, keeping pace with the speed of bacterial evolution itself.
In laboratories where the boundary between biology and computation has grown thin, scientists have done something quietly extraordinary: they asked an artificial intelligence to imagine viruses that nature never made, and the viruses worked. Published in Science, the research describes how AI-designed bacteriophages — living medicines conjured from learned patterns in DNA — defeated antibiotic-resistant bacteria that natural viruses could not overcome. It is an early signal that medicine may soon be able to design its own evolutionary answers to one of its oldest and most dangerous adversaries.
The scenario sounds like science fiction: artificial intelligence inventing viruses that have never existed, then deploying them as living weapons against bacteria immune to every antibiotic we have. But the science is real, and it may represent one of medicine's most important recent advances.
Researchers published their findings in Science, describing how they set out to design bacteriophages — viruses that infect only bacteria — as a new class of medicine. The challenge was formidable. Building a functional viral genome from scratch requires not just genes, but regulatory elements and countless precise interactions between components. Predicting protein structures is hard enough; writing an entirely new instruction manual for a living organism is something else entirely.
The team trained an AI system called Evo on millions of DNA sequences from ΦX174, one of biology's most studied bacteriophages. The model absorbed enough examples to learn the underlying rules of viral construction, then generated approximately 700,000 novel genomes — designs nature had never assembled. From these, 285 candidates were chemically synthesized and inserted into E. coli cells. Sixteen produced fully functional viruses, complete with new gene combinations the biological world had never seen.
The decisive test came when bacteria were allowed to evolve resistance to the natural ΦX174 virus, mirroring what happens clinically when infections become drug-resistant. The researchers responded not with a single phage, but with a cocktail of all 16 AI-designed variants. The resistant bacteria were rapidly overwhelmed. A comparable cocktail of naturally occurring relatives proved far less effective.
The elegance of the strategy lies in its diversity: each AI-designed phage attacked bacteria differently, making it nearly impossible for any single mutation to defend against all of them at once. As antibiotic resistance grows into one of medicine's most intractable crises, this research suggests a new path — designing phages on demand, tailored to specific infections, potentially compressing years of development into weeks.
The scenario reads like speculative fiction: scientists use artificial intelligence to invent viruses that have never existed in nature, then deploy them as living weapons against bacteria that have grown immune to every antibiotic we have. But the science underneath is real, and it may point toward one of medicine's most urgent needs.
Researchers published their findings in Science earlier this year, describing how they designed bacteriophages—viruses that infect bacteria and nothing else—with the goal of turning them into a new class of medicine. The challenge they faced was not small. While AI has proven useful for predicting protein structures or analyzing stretches of DNA, building an entire functional genome from scratch is orders of magnitude harder. A working virus requires not just genes, but control sequences, regulatory elements, and countless interactions between components, all of which must function together in precise coordination.
The team started by training an AI system called Evo on millions of DNA sequences from a tiny bacteriophage known as ΦX174, one of the most thoroughly studied viruses in biology. This virus naturally infects only E. coli bacteria. The researchers fed the AI enough examples that it began to learn the underlying rules of how such viruses are constructed. The model then generated approximately 700,000 completely novel viral genomes—designs that had never appeared in nature. From this vast library, the scientists selected 285 candidates that looked most promising and had the DNA chemically synthesized in the laboratory.
When those synthetic genomes were inserted into E. coli cells, something remarkable happened: 16 of them produced fully functional viruses. These were not minor variations on existing phages. They contained new combinations of genes and DNA sequences that nature had never assembled. Yet despite their novelty, they still managed to infect bacteria successfully. The researchers had essentially written new chapters in the viral instruction manual and the cells had read them.
The real test came next. The team allowed E. coli bacteria to evolve resistance to the natural ΦX174 virus, mimicking what happens in the clinic when bacteria develop immunity to a drug. Normally, once that resistance takes hold, treatment fails. But instead of relying on a single virus, the researchers mixed all 16 AI-designed phages into a cocktail. The bacteria that had become resistant to the natural virus were rapidly overwhelmed by the mixture. A comparable cocktail made from naturally occurring relatives of ΦX174 proved far less effective.
The reason reveals something elegant about the strategy. Each of the AI-designed viruses attacked the bacteria in slightly different ways. A genetic mutation that shields bacteria from one phage does not necessarily protect it from another. By varying the attack vectors, the cocktail made it nearly impossible for the bacteria to evolve a single defense that would work against all of them simultaneously.
This matters because antibiotic resistance has become one of modern medicine's most intractable problems. Infections like tuberculosis and MRSA are growing harder to treat as bacteria evolve immunity to the drugs we have. Phage therapy—using viruses to kill bacteria—has long been proposed as an alternative, but finding suitable phages in nature is slow and unpredictable. This research suggests a different path: designing new phages on demand, tailored to attack whatever bacterial infection needs treating. If that vision becomes reality, it could compress years of development into weeks, and give medicine a tool that evolves as fast as the bacteria it fights.
Citations marquantes
Rather than searching nature for suitable phages, we may one day be able to design entirely new ones on demand, tailored to attack specific bacterial infections— Researchers, as reported in Science
Envie de l'histoire complète ?
Lire l'original sur Newstalk ZB ↗
Science & Space