A petri dish filled with dead bacteria isn’t typically a reason to celebrate—but for Stanford researcher Brian Hie, it marked a breakthrough in his quest to engineer synthetic life.
The killer was a bacteriophage, a virus that targets and destroys bacteria while leaving human cells unharmed. These viruses have evolved over millions of years to eliminate harmful bacteria and are seen as powerful allies in combating antibiotic resistance.
But in this case, evolution wasn’t the designer—an AI system similar to ChatGPT wrote the entire viral genome. This artificial genetic blueprint enabled the synthetic virus to replicate, infect, and kill bacteria, marking a major milestone toward AI-created life forms.
Synthetic but Not Truly Alive
Despite behaving like natural viruses, these engineered versions aren’t technically “alive.” Viruses consist of minimal genetic material and require a host—such as bacteria—to reproduce and spread.
Still, they represent the closest scientists have come to crafting new life using generative AI. The breakthrough could lead to new treatments for dangerous infections and offer insights into constructing more complex artificial cells.
“This is the first time AI systems are able to write coherent genome-scale sequences,” Hie told Nature. The findings are currently available as a preprint on bioRxiv and have yet to undergo peer review.
The genetic foundation of all life on Earth is surprisingly straightforward. Just four molecules—A, T, C, and G—form three-letter combinations that code for amino acids and proteins.
Rewriting DNA for Real-World Benefits
Synthetic biologists tweak this code by inserting helpful genes or removing harmful ones. Thanks to their efforts, bacteria like E. coli—a staple in laboratories—can now produce insulin and other medications.
Now, generative AI is pushing the field even further.
These models can already invent new DNA sequences, proteins, and complex molecular structures from scratch. But constructing an entire functional genome is far more challenging—it has to encode the machinery of life and ensure all the components work together properly.
“Many essential biological functions don’t come from individual genes, but from intricate interactions across whole genomes,” the researchers noted.
To tackle this, the team used Evo 1 and Evo 2—AI models developed by the nonprofit Arc Institute. Unlike typical language models trained on internet text, Evo 2 was trained on about 128,000 genomes containing 9.3 trillion DNA base pairs from across all domains of life, making it the largest AI model in biology to date.
The models learned how DNA changes affect RNA, proteins, and overall cell function, enabling them to generate entirely new proteins and compact genomes.
From Digital Designs to Biological Blueprints
Evo 1, for instance, designed novel CRISPR gene-editing tools and bacterial genomes—but many of the latter included unnatural sequences that failed to function in living bacteria. Evo 2 went further, creating a full human mitochondrial genome that produced natural-like proteins, along with a minimal bacterial genome and a yeast chromosome. However, none of these designs have yet been tested in living cells.
The latest research zeroed in on a more basic biological target: bacteriophages. These viruses specifically infect bacteria and are already being tested in clinical trials as a weapon against antibiotic-resistant infections. In principle, engineered versions could be even more potent.
The researchers started with phiX174, a simple bacteriophage consisting of a single DNA strand with 11 genes and 7 regulatory regions. Despite its small genome, it has everything needed to infect, replicate, and spread. PhiX174 is well-known in synthetic biology—its genome has already been fully sequenced and lab-synthesized, making it ideal for genetic experimentation. It’s also considered safe and “has consistently served as a key model in molecular biology,” the team noted.
Although the Evo AI models had already been trained on roughly two million genomes, the researchers further refined them with a specialized “masterclass” in phage DNA. They introduced biological constraints found in phage genomes and added prompts to promote creative variations.
The AI systems then produced thousands of genome candidates, some with clear flaws. While many drew from familiar patterns in their training data, others generated entirely novel genetic sequences. On average, about 40% of the DNA overlapped with phiX174, but the rest represented completely new structures.
Testing AI-Built Viruses in the Lab
From these, the researchers selected and synthesized 302 candidates to test their bacteria-killing potential. Sixteen of the AI-designed genomes behaved like functional bacteriophages: they infected E. coli, replicated inside the cells, ruptured them, and spread to surrounding bacteria. Unexpectedly, combinations of these synthetic phages were even able to attack other E. coli strains they weren’t specifically engineered for.
“These findings show that genome-scale language models can design working phage genomes,” the team concluded.
Generative AI could greatly speed up the creation of synthetic life. Instead of relying on slow trial-and-error, models like Evo have already learned how genes and molecules interact.
With more testing, this could transform phage therapy and help treat antibiotic-resistant infections in humans and crops.
However, AI-designed viruses raise safety concerns. To mitigate risks, researchers excluded human-infecting viruses from Evo’s training data, used supervised guidance, and worked only with well-studied, safe organisms like phiX174 and E. coli.
Experts still warn that similar tools could be misused to enhance harmful viruses. Designing larger genomes, such as those of bacteria, is also far more complex, and ethical debates around synthetic life are growing.
Even so, the team sees this as an early step toward AI-built living systems. With more progress, Hie says, “the next step is AI-generated life.”
Read the original article on: Singularity Hub
Read more: DeepMind Unveils its First Thinking Robot AI
Test message