AI Just Learned to Make Viruses From Scratch

William Wei
7 days ago
5 min read

What?

In February 2025 the Evo 2 underlying genome AI model at Stanford was released as a preprint. Now, 2 days ago, the scientists at Stanford officially used AI to design completely new virus genomes from scratch. With a virus’s genome essentially its instruction manual, scientists typically study existing viruses and modify their genetic code. The AI’s prompt was essentially, “Based on everything you know about genetic code, what could a functional virus genome look like?” The AI generated new genetic sequences, and scientists tested some of the designs in the lab. Out of 302 generated designs, the researchers synthesized the most promising ones. 16 of them actually worked as bacteriophages.

Why is this a big deal?

The important part isn’t really the fact that they made some new viruses. It’s the fact that AI successfully designed an entirely new functional viral genome. Previously, AI has been used to help design individual biological components, like proteins or potential antibiotics. However, this is a much larger step. AI generated the equivalent of an entire set of genetic instructions, with some of those instructions actually working when tested in the lab. Researchers are calling this a major turning point in synthetic biology. The technology behind the breakthrough is based on the same technology as ChatGPT. These large language models learn patterns in large collections of text and use those patterns to predict what should come next. Genome language models apply a similar concept to DNA sequences. Instead of learning relationships between words, they learn the patterns in genetic language shared across genomes.

Evo 1 and Evo 2 were trained on large collections of genetic sequences. Researchers then used the models to generate potential bacteriophage genomes based on the bacteriophage ΦX174.

The important part is that the AI was not simply copying an existing virus. The researchers were testing whether a model could generate a complete genome with the many interacting genetic elements necessary for a virus to function. This is way more difficult than designing one gene. A genome is an interconnected system in which one part affects how other parts function when changed. The researchers needed to determine whether the AI-generated sequences would actually function when synthesized and tested.

They did, by the way.

The researchers synthesized hundreds of the most promising AI-designed phages and tested them against E. coli. The scientific study also reported that some of the generated phages outperformed the original ΦX174 in laboratory measurements. A combination of generated phages overcame the resistance E. coli had developed to the original phage.

Bacteriophages?

The choice of bacteriophages was important. Bacteriophages are viruses that specifically infect bacteria. Unlike viruses that infect humans, these phages do not infect human cells. Scientists have been studying them as a possible way to fight bacterial infections. Specifically, infections that are caused by bacteria that have become resistant to antibiotics. With antimicrobial resistance a major global health problem, this new synthesis of bacteriophages could offer a potential solution.

According to the World Health Organization, approximately one in six lab-confirmed bacterial infections worldwide was resistant to antibiotic treatment in 2023. Between 2018 and 2023, resistance increased in more than 40% of the pathogen-antibiotic combinations monitored by WHO.

Traditional antibiotics work by attacking bacteria in particular ways. But bacteria evolve, and some eventually develop mechanisms that allow them to survive those drugs. Phages offer a different approach than relying on a chemical antibiotic. Scientists can potentially use a virus that specifically targets specific bacteria. The WHO describes phage therapy as a promising approach for combating antimicrobial resistance, but also emphasizes that additional clinical evidence is needed before phages can become widely available as routine human treatments. AI could eventually make this approach more powerful and available. Instead of searching through naturally occurring phages and hoping to find one that attacks the specific bacteria, scientists can just use AI to help design phages with the desired characteristics (assuming the best).

This creates the possibility of a future in which biological treatments are designed much more rapidly and precisely.

AI “writing” Biology?

The most important part of this research may not be the 16 viruses. It’s the demonstration that an AI model can produce a complete biological system that works in the real world. For decades, biology has largely been a science of observation and modification. That means scientists study existing organisms and identify useful genes to modify biological systems based on what nature has laid out.

Generative AI throws all of that out the window.

If AI models can learn enough of the patterns that govern biological systems, they may eventually help scientists design systems that do not already exist in nature. This could even extend beyond phages, with researchers having suggested similar approaches that could potentially be applied to biological tools like enzymes and gene therapies. However, the current experiment is still limited. A small bacteriophage is way simpler than a living organism such as a bacterium, plant, or even animal.

Uh oh?

The same capabilities that make technology scientifically exciting also create a pretty big problem. If AI can generate functional viral genomes, scientists and policymakers have to consider how the technology could be used in the future.

The current study was deliberately designed around bacteriophages rather than viruses that threaten humans. The researchers took steps to limit biological risk, including restricting the training data used for the work and conducting the experiments in a secure lab. The viruses produced in this study were designed to target bacteria and don’t actually pose a threat to people (hooray!).

Nevertheless, biosecurity experts have argued that the development raises urgent questions about how AI-assisted biological design should be governed and handled. The concern is not that the viruses already created could be dangerous to humans. It actually demonstrates that AI can now produce functional viral genomes. This creates a capability that will need to be carefully monitored as future models become stronger.

A commentary from researchers at John Hopkins’ Center for Health and Security argues that the question is increasingly shifting from whether AI can design viral genomes to how society can guarantee that the capability is used safely and correctly.

A new era?

It’s a little too early to say that AI can just “make life”. Viruses are in a weird position in biology because of their simplicity relative to living cells. The bacteriophage genomes used in this research were only in the thousands of DNA base pairs in terms of length. Compared to the genomes of relatively simple cellular organisms, they are drastically smaller. Still, the experiment marks a large change in what scientists can do with AI. It’s no longer just used to read and interpret; it can now help scientists generate new biological data.

The immediate application could be medical. Designing new bacteriophages to attack bacteria that have become resistant to existing treatments could be extremely useful in helping this global health issue. However, the long-term implications are a lot broader. It could potentially change how researchers develop medicines, study evolution, and engineer any biological systems.

The breakthrough also demonstrates why scientific progress and safety have to advance relative to one another. As AI becomes stronger and more capable of designing biology, researchers will have to answer a few questions. First, “what can we make?” and second, “what should we make?” and finally, “how do we make sure it’s used correctly?”

The 16 AI-designed phages are a little more than just a collection of new viruses. They are evidence of a huge transition in science. We used to use computers to study and understand biology. I guess it’s time to use those computers to design new parts of it now.