AI-Designed Viruses Mark New Biology Breakthrough, Raise Safety Concerns

Researchers have demonstrated that artificial intelligence can be used to design previously unknown viral genomes, marking a significant development in synthetic biology while also raising questions about the future security of AI-enabled biological research.

The work, led by researchers including Stanford University computational biologist Brian Hie, involved AI-designed viruses known as bacteriophages, which infect bacteria rather than humans. The researchers used genome-focused AI models to generate novel viral designs and then tested selected candidates in laboratory experiments. The study reports that 16 AI-generated phages were viable.

The development has potential applications in medicine, particularly in the search for new ways to tackle antibiotic-resistant bacterial infections. At the same time, biosecurity experts have warned that increasingly capable biological AI systems could eventually create new risks if the technology is misused or developed without appropriate safeguards.

AI-designed viruses mark a new step in synthetic biology

Artificial intelligence has increasingly moved beyond text, images and software into biological research.

Instead of learning patterns from written language, genome language models analyse genetic sequences and learn relationships within biological data. Researchers can then use those models to generate new sequences based on patterns learned from existing genomes.

Stanford’s Brian Hie has been working at the intersection of artificial intelligence and biology. His research includes genome-scale modelling and design using systems such as Evo, an AI model developed to work with genetic sequences. Stanford describes Evo as a generative model capable of writing genetic code and exploring biological systems at genome scale.

The latest work extends that approach from individual genes and proteins toward complete viral genomes.

The researchers reported that their models were able to generate functioning bacteriophage genomes that were sufficiently different from naturally occurring sequences to demonstrate substantial evolutionary novelty.

The viruses target bacteria, not humans

Despite headlines about AI creating new viruses, an important distinction is that the viruses produced in this research are bacteriophages.

Bacteriophages are viruses that infect bacteria. They are fundamentally different from viruses that cause human diseases.

The researchers focused their work on bacterial targets and designed the experiment around bacteriophage biology. The resulting viruses were tested for their ability to infect bacterial cells in controlled laboratory settings.

The study therefore does not demonstrate that AI has created a virus capable of infecting humans.

That distinction is important when assessing both the scientific achievement and the potential risks. Current research is focused on bacterial viruses, while the possibility of designing viruses affecting humans remains a much more difficult and uncertain problem.

Researchers see potential against drug-resistant infections

One of the most promising applications of the technology is phage therapy.

Phage therapy uses bacteriophages to target bacteria and has attracted renewed scientific interest as antibiotic resistance makes some infections increasingly difficult to treat.

The researchers reported that some of the AI-generated phages performed strongly against E. coli in laboratory experiments. They also found that a combination of generated phages could overcome resistance in several bacterial strains in their experiments.

The potential significance is that AI could eventually help researchers search biological sequence space more efficiently than conventional trial-and-error approaches.

Rather than relying only on viruses already discovered in nature, scientists could potentially use computational models to explore new biological designs and then test promising candidates under controlled conditions.

However, the research is still far from establishing an AI-designed treatment that can routinely be used in patients.

From DNA data to biological design

The concept behind genome language models is similar in principle to the way language models identify patterns in text.

Instead of learning relationships between words, genome models learn patterns within DNA sequences.

Stanford researchers have previously demonstrated that generative models can produce novel genetic sequences with biological functions. Earlier work involving Hie and colleagues showed that genomic language models could be used to design functional genes that did not have significant similarity to known natural proteins. That research was published in Nature.

The bacteriophage research takes the concept further by demonstrating that an AI-generated sequence can form part of an entire functioning viral genome.

Researchers described this as an important step toward genome-scale biological design.

Why the breakthrough is raising biosecurity questions

The same capability that could help researchers develop new medical tools also creates a potential dual-use problem.

Tom Inglesby, director of the Johns Hopkins Center for Health Security, was not involved in the research but co-authored commentary on the work, according to the NPR report supplied for this article.

Inglesby argued that the scientific achievement has potential benefits but warned that governance and safeguards have not necessarily developed at the same pace as the technology.

The concern is not that the viruses created in this particular experiment are human pathogens. Rather, experts are considering what could happen if increasingly powerful biological design systems become widely accessible and are eventually applied to organisms with greater potential to cause harm.

Experts call for stronger oversight

Kevin Esvelt, a genetic engineer at MIT who studies biological risks, has also argued that regulation needs to keep pace with advances in biological design.

One area of policy discussion involves companies that synthesise custom DNA sequences.

The supplied NPR report notes that a bipartisan US Senate bill has been proposed that would require companies selling custom DNA sequences to screen orders for potentially dangerous sequences.

Such screening is one part of a broader biosecurity approach that could involve research oversight, laboratory safeguards, sequence screening and responsible access to advanced biological design systems.

The precise regulatory framework for AI-enabled biology remains under development.

AI does not eliminate the difficulty of biological research

The development should also be viewed in the context of the significant gap between generating a genetic sequence on a computer and successfully producing a functioning biological system.

Most computationally generated biological designs do not necessarily work when tested experimentally.

The researchers’ own results illustrate this challenge: although many candidate sequences could be generated computationally, only a subset of the designs tested experimentally produced viable bacteriophages. The published research reports 16 viable phages from the experimental testing.

This means AI has not removed the need for laboratory research, scientific validation or safety controls.

Instead, the technology gives researchers a new way to explore possible biological designs before selecting candidates for further investigation.

What this could mean for medicine

If the technology continues to develop safely, AI-assisted genome design could eventually contribute to several areas of biotechnology.

Potential applications include designing new biological tools, studying how microbes evolve and developing targeted approaches against difficult bacterial infections.

The research team has highlighted the possibility of using generative biology to respond more quickly to biological threats.

For example, if a bacterial pathogen develops resistance to existing treatments, researchers could potentially use computational methods to explore new bacteriophage candidates rather than relying exclusively on naturally occurring phages.

However, translating laboratory findings into human medical treatments requires extensive additional research, including safety testing, clinical studies and regulatory approval.

The current study does not establish that AI-designed viruses are ready for clinical use.

A new challenge for AI governance

The research also adds to a growing debate about how governments should regulate AI systems capable of interacting with biological science.

AI development has traditionally focused on software, information and digital systems. Generative biology introduces a different category of risk because computational outputs can potentially be translated into physical biological systems.

That makes safeguards particularly important.

Researchers and policymakers face the challenge of encouraging legitimate scientific applications while reducing the possibility that powerful biological design tools could be misused.

The issue is complicated further because biological research is increasingly global, while regulation is generally organised at the national level.

AI and biology enter a new phase

The development of AI-designed bacteriophages represents a significant step in the growing relationship between artificial intelligence and biology.

The research demonstrates that genome language models can move beyond analysing biological information and help generate novel, functioning genetic systems.

For medicine, the technology could eventually contribute to new approaches for difficult bacterial infections and other biological challenges.

For biosecurity experts, however, it is also a reminder that biological capabilities are becoming increasingly programmable.

The immediate research involved bacteriophages that target bacteria, not viruses designed to infect humans. But the broader lesson is that AI is becoming capable of participating in increasingly complex biological design.

As these systems improve, scientists, governments and technology companies will face growing pressure to ensure that the benefits of AI-enabled biology advance alongside effective safety and security measures.