NEW YORK, Sept 20: Artificial intelligence, known for drawing cat pictures and writing emails, has now achieved a major scientific milestone by composing working genomes. A research team in California has used AI to design new genetic codes for viruses, successfully creating several viruses that replicated and killed bacteria.

Scientists from Stanford University and the nonprofit Arc Institute in Palo Alto describe these AI-generated viruses as “the first generative design of complete genomes.” Their findings, published in a preprint paper, hold promise for developing new treatments and speeding up research on engineered cells. Jef Boeke, a biologist at NYU Langone Health who reviewed the study, called it an “impressive first step” toward AI-designed life forms, noting the AI’s surprising ability to create viruses with new genes, shortened genes, and novel gene arrangements.

However, this is not yet true AI-designed life, since viruses are not considered alive. They are simple strands of genetic material with relatively small genomes.

The study taps into AI’s expanding role in automating drug development, which could make pharmaceuticals faster and cheaper to produce. The researchers focused on variants of a bacteriophage virus called phiX174, which infects bacteria and has just 11 genes with about 5,000 DNA letters.

They trained two versions of an AI called Evo, which operates similarly to large language models like ChatGPT, but instead of text, Evo was trained on the genomes of around 2 million bacteriophage viruses.

To test the AI-generated genomes, the team chemically synthesized 302 designs and introduced them to E. coli bacteria. The breakthrough came when the scientists observed plaques of dead bacteria on petri dishes—a clear sign the AI-designed viruses were replicating and destroying the bacteria. Microscope images showed viral particles resembling fuzzy dots.

Brian Hie, who leads the lab at the Arc Institute where this research was conducted, described the experience as striking, highlighting the tangible reality of “AI-generated spheres.”

Out of 302 genome designs, 16 were successful, with the AI-created bacteriophages replicating and lysing the bacteria.

J. Craig Venter, a pioneer who created some of the first organisms with lab-made DNA nearly 20 years ago, likened the AI method to a faster version of trial-and-error experimentation. His 2008 bacterium with a synthetic genome was produced through lengthy testing of various genes, which he calls a “manual AI version.”

The key advantage of AI is speed. The technology already earned a Nobel Prize in 2024 for predicting protein structures, and investors are pouring billions into AI-driven drug discovery. Recently, Boston company Lila raised $235 million to develop AI-operated automated labs.

Computer-designed viruses also have commercial potential. Doctors have explored “phage therapy” to treat bacterial infections, and similar approaches are being tested to protect cabbage from black rot, a bacterial disease.

Samuel King, the student leading this project in Brian Hie’s lab, said the technology holds great promise. Since many gene therapies use viruses as vehicles to deliver genes into patients, AI could help create more effective viral carriers.

The Stanford team deliberately avoided training their AI on viruses that infect humans, but they warn that the technology poses risks. Other scientists, motivated by curiosity, good intentions, or malice, could potentially use the methods to engineer dangerous human pathogens.

J. Craig Venter urged caution, especially regarding random viral enhancement research where outcomes are unpredictable. “If someone did this with smallpox or anthrax, I would have grave concerns,” he said.

Whether AI can design genomes for larger, more complex organisms remains uncertain. For example, E. coli’s genome is roughly 1,000 times larger than phiX174’s, making the complexity immense. Boeke notes this complexity could surpass even the number of subatomic particles in the universe.

Testing AI-designed genomes for larger organisms is also difficult. While some viruses can “boot up” from synthesized DNA strands, this is not true for bacteria, mammals, or humans. Instead, scientists must gradually edit existing cells using genetic engineering, which remains a slow and complex process.

Despite these challenges, Jason Kelly, CEO of Boston-based cell-engineering company Ginkgo Bioworks, argues that pursuing this effort is essential. He envisions automated labs where AI proposes genomes, tests them, and uses the results to refine further designs.

Kelly called such a development “a nation-scale scientific milestone,” emphasizing that cells are the fundamental units of life. “The US should make sure we get to it first,” he said.