Science & Technology24 August 20264 min read
AI models design working bacteriophage genomes, 16 of 285 infect E. coli
Researchers from Stanford University and the Arc Institute have used the generative AI models Evo 1 and Evo 2 to design complete bacteriophage genomes that work in the laboratory. Of 285 AI-designed genomes that were physically synthesised and tested, 16 produced viable phages that infected E. coli.
Some of the designs overcame bacterial resistance that had defeated the original virus. One combined a viral protein with other genetic changes in a way that conventional engineering had struggled to achieve. The AI could explore gene combinations that would be extremely difficult for humans to test one by one.
Why phages matter
Bacteriophages, or "bacteria eaters", are viruses that selectively target and kill bacteria. They have been studied for more than a century and do not infect human cells. The ΦX174 phage was the first complete DNA genome to be sequenced, in 1977.
Antimicrobial resistance (AMR) is one of the top global public health and development threats. That has revived interest in phage therapy, an older approach to treating infections, because phages can destroy bacteria that resist antibiotics. The approach has limits: phages are highly specific, so one effective against a bacterial strain may not work against another, and bacteria can develop resistance to phages too.
The experiment falls within generative biology, where models learn the patterns of DNA, RNA and proteins to design new biological entities from scratch. The progression runs from reading genomes in the 1970s, as with ΦX174, to writing them by chemical synthesis in the 2000s, modifying them through gain-of-function work and targeted edits in the 2010s, and now designing them with AI.
The potential applications listed are:
- customised phage therapies against multidrug-resistant "superbugs";
- engineered viral vectors, such as Adeno-Associated Viruses, to deliver CRISPR gene-editing payloads to specific organs;
- faster design of antigens, monoclonal antibodies and therapeutic proteins against emerging pathogens;
- oncolytic viruses that attack cancer cells while sparing healthy tissue.
Biosecurity and governance concerns
The central worry is dual-use risk. The capabilities that support phage therapy, vaccines and gene therapy could also be misused to design harmful agents, potentially lowering barriers to bioterrorism. The concern is that AI may amplify the abilities of people or laboratories that already have biological expertise and infrastructure, not that AI would independently create a dangerous pathogen today.
An AI-designed virus released intentionally or by accident could interact unpredictably with ecosystems, potentially jumping species barriers. Faster design cycles add to the concern: future systems could cut the time spent on literature review, modelling and experimental design and, combined with automated laboratories, shorten the gap between design and testing.
Current safeguards also have gaps. Traditional DNA-synthesis screening checks whether an ordered sequence resembles a known pathogen or toxin, but AI-generated designs may be novel and resemble nothing on the list. Patent law is unsettled as well: global intellectual property rules are unprepared for AI-generated life forms, and it is unclear who would hold a patent on a synthetic virus engineered entirely by a machine learning algorithm.
Measures proposed
The steps suggested include:
- mandatory DNA-synthesis screening, with Know Your Customer (KYC) norms for commercial gene-synthesis providers;
- function-based biosecurity frameworks that assess the likely biological effects of a design, extended across synthesis providers, laboratories, databases, computing infrastructure and institutional biosafety mechanisms;
- updating the 1972 Biological Weapons Convention to cover AI-generated pathogens and synthetic biology;
- oversight of the computing power used to train frontier biological models, as with nuclear materials;
- red teaming of biological models by developers, to ensure they refuse prompts to generate harmful pathogens.
For India, AI-driven biological design bears on drug discovery, genomics, vaccine development and AMR research. The country needs to strengthen biomedical AI, computing, biological datasets and biosecurity through initiatives such as the IndiaAI Mission, while balancing AI sovereignty with strong biosafety safeguards.