AI designed a virus, a scientist activated it — and here the story gets complicated
Researchers from Stanford and the Arc Institute used artificial intelligence to design the genetic code of bacteriophages — viruses that attack bacteria — and created active viruses from them in the lab. This achievement could help in developing treatments against antibiotic-resistant bacteria, but experts warn that the improving ability of AI to design biological systems raises urgent questions of safety and oversight.
Until recently, the discussion around artificial intelligence focused mainly on what it is capable of writing, drawing, or calculating. Now it turns out that the question is much broader: can it also design things that will function in the biological world? A team of researchers from Stanford and the Arc Institute managed to use artificial intelligence to design the genetic code of bacteriophages — viruses that attack bacteria — and then create active viruses in the lab based on those designs.
In a study published last week in the journal Science, artificial intelligence models created hundreds of new phage designs. Out of 285 designs tested in the lab, 16 led to the creation of active viruses that were able to multiply and attack E. coli bacteria. Later, the researchers showed that a combination of some of the new phages was able to also harm bacteria that had developed resistance to the original phage.
To understand how artificial intelligence can even design a virus, one needs to think of DNA as a language. Prof. Ran Nir-Paz, an expert in infectious diseases at Hadassah Hospital and the clinical lead of the Israeli Phage Therapy Center, explains that the researchers developed models trained on large amounts of genetic data that learned to identify the rules and patterns by which genomes are built and function.
"Just as there is Claude and Gemini, which are capable of using ordinary language to produce sentences, paragraphs, or even books, the researchers developed two models, Evo 1 and Evo 2, whose role is to use the language of DNA and RNA," he explains.
In the current study, this analogy takes on real meaning: if language models produce texts of different lengths, here the models produce genetic sequences of different lengths and levels of complexity. "From short words and sentences to long books," explains Prof. Nir-Paz. "A long book would be a monkey, a human, or a dog, while the paragraphs or short sentences are supposed to be bacteriophages, the subject of this article, viruses of bacteria."
When the computer starts writing biology
The idea of designing biological systems on a computer is not new. According to Prof. Nir-Paz, one of the pioneers in the field was the American geneticist Craig Venter, one of the leaders of the Human Genome Project. About a decade ago, long before the emergence of generative AI, a breakthrough occurred with the creation of a bacterium with a minimal genome designed on a computer and built in the lab so that it could divide and multiply.
"Its goal was quite similar to what the researchers did here: to use existing knowledge to create a life form capable of dividing on its own, and to endow it with various properties," explains Prof. Nir-Paz. What has changed since then? Mainly the power of the computational tools. Researchers used technologies resembling those behind language models to learn from huge amounts of genetic data.
In the eyes of Prof. Nir-Paz, the most significant contribution of such tools may be precisely in deepening our understanding of biology. "In the genome of bacteriophages, we know the role of about 40% of the genes and proteins. We understand less of everything else." Models that know how to identify patterns within huge amounts of genetic data may help to decipher those parts that until today remained a mystery.
The dark side of synthetic biology
This fear is currently at the heart of the discussion. As models become better at designing biological systems, so does the potential to use them not only to understand nature or develop new treatments, but also to create systems with properties that did not exist before. From here arises a question: what happens when this ability is directed in dangerous directions, or when it simply goes out of control?
Prof. Tal Brosh, director of the Infectious Diseases Unit at Assuta Ashdod Hospital, emphasizes that in the case of the current study, there is no direct danger to humans, as bacteriophages attack bacteria and not human cells. However, he warns:
"There could be misuse or a laboratory accident that ends in them taking a virus, a bacterium, or some creature, and causing it to be more violent, more contagious, or resistant to treatments and vaccines."
To illustrate how much the possibility of recreating viruses is not just a theoretical idea, he mentions smallpox. A few years ago, a group of researchers managed to recreate in the lab a horsepox virus, relying on its genetic sequence from the internet. "This was done using technology much less advanced than that which exists today," describes Prof. Brosh.
The fear is that artificial intelligence will be able in the future to expand these capabilities, and not only restore an existing virus but also assist in the search for changes that will affect its properties. This raises a complex question of boundaries and oversight. On one hand, you don't want to place barriers before science, but on the other, you don't want everything that a person thinks of to be possible to execute.
Despite the risks, the same tools could grant medicine capabilities it did not have in the past. For example, they could help better predict mutations in influenza or coronavirus strains and design vaccines that protect against a wider variety of viruses. "This is a bigger story hiding behind the new study. As this ability improves, so does the responsibility to ensure that it remains under control," concludes Prof. Brosh.
