AI designs 16 viable viruses from scratch

Researchers at Stanford University and the Arc Institute have trained artificial intelligence to interpret the “grammar” of genetic material and design 16 new viable viruses from scratch, ERR reported. The achievement could eventually help treat inherited diseases, but it may also make biological threats more difficult to control.

Large language models can write poems and essays because they learn from vast amounts of text which words commonly appear near one another. A new model called Evo 2 applies a similar approach to DNA, arranging the building blocks of genetic material instead of words.

DNA consists of four nucleotides and follows rules that resemble a more complex biological alphabet. Because the model can retain far more information at once than the human mind, it can identify patterns that researchers may previously have overlooked.

To find these underlying rules, the researchers first fed Evo 2 nine trillion nucleotides from DNA sequences belonging to animals, plants and microbes. The algorithm began identifying recurring evolutionary patterns and soon learned to generate genes that encode proteins.

Because the genomes of complex organisms contain billions of coded letters, the researchers initially focused on viruses, whose genomes often contain only a few thousand nucleotides. Study co-author and Stanford University graduate student Samuel King said designing viruses seemed like a logical first step.

The team then fine-tuned the model using one specific virus, Phi X-174, and the DNA sequences of more than 14,000 similar microorganisms. The bacteriophage, a virus that infects bacteria, attacks only intestinal bacteria and leaves human cells unaffected. The researchers hoped this would reduce the risk of creating pathogens capable of infecting other organisms.

The artificial intelligence generated hundreds of thousands of potential new viral genomes. Researchers selected 285 promising variants and tested them in Petri dishes by inserting synthetic DNA sequences into bacteria and observing the results.

Most of the synthetic sequences had no effect on bacterial growth. In one culture, however, transparent spots began appearing on the surface of the cells, a clear sign that an artificial virus was viable. By the end of the experiments, the biologists had created 16 entirely new viruses. Some of the machine-designed genomes differed considerably from their natural counterparts and contained unique solutions.