For the first time, artificial intelligence designs new viruses

A research team at the Arc Institute in Palo Alto, California, has generated new types of virus using artificial intelligence. It is the first time this has been done, according to the study published in the journal Science and reported by the New York Times.

The tool used is a model called Evo, whose workings are compared to those of ChatGPT. Instead of words, the model works on the four letters of the DNA alphabet: A, C, G and T.

Training was carried out on genetic sequences from millions of organisms, amounting to around 9,000 billion nucleotides. The model then produced new DNA molecules, which the researchers inserted into bacterial cells.

The limits set upstream

The natural reference point is Phi X-174, a virus that occurs in nature and infects only bacteria. The viruses obtained behave the same way: they infect bacteria and pose no threat to human beings.

That choice was not accidental. Viruses that infect humans, and comparable pathogens affecting other organisms, were excluded from the training data.

The computational biologist Brian Hie, of Stanford University, explained the reason for the decision. «Volevamo semplicemente essere estremamente cauti» (We simply wanted to be extremely cautious), he said.

A regulatory question remains open. The United States has introduced rules limiting high-risk research in the life sciences, but computational research of this kind does not appear to fall under those provisions.

This is the gap between the scope of the rules and the scope of the available techniques. The rules were written with the laboratory in mind, while part of the work now takes place on models trained on the genome.

The results show that a model can produce working genomes from known sequences. Which categories of organism remain outside its reach depends, for now, on whoever selects the starting data.