Much of the recent application of artificial intelligence in computational biology has centered on protein design. Because proteins direct biochemical reactions and structure living cells, generating custom proteins offers a direct method for introducing novel biological functions. However, because the genetic code acts as a layer of abstraction between underlying DNA and resulting proteins, it was not initially obvious what AI architectures trained directly on raw genomic sequences could accomplish.
Despite that uncertainty, researchers built large genome models to analyze DNA directly. Early tests showed these systems could output DNA sequences that successfully encoded functional proteins within bacteria while accurately mimicking the gene structures found in complex cells. Taking the technology a step further, a team based at Stanford University has now employed these models to generate complete viral genomes designed to infect bacteria.
The resulting viral constructs are not entirely fictional creations built from scratch. All of the synthetic viruses generated by the AI models remain closely related to known, pre-existing viruses. Nevertheless, the model-designed genomes incorporate distinct structural and sequence features that would be exceptionally challenging to develop through standard evolutionary pathways.
What it means
By demonstrating that large genome models can construct functional viral code, the study marks a significant shift in AI-driven genetic engineering. While the current research focused strictly on viruses that target bacterial hosts, the Stanford team raised important considerations regarding future capabilities. They cautioned that the research community should begin preparing for the eventual development of related AI systems capable of designing synthetic viruses tailored to infect vertebrates.
