Stanford Uses Large Genome Models to Design Novel Bacteriophages

👁 119 reads

Researchers at Stanford University have used large genome AI models - Evo 1 and Evo 2 - to generate the complete genomes of viruses that infect bacteria, marking a step beyond earlier work that focused on designing individual proteins or gene clusters.

The test case was ΦX174, a well-characterized bacteriophage that infects E. coli and belongs to a broader family of bacterial viruses. According to the researchers, every virus the models produced is closely related to that existing virus, so the work does not involve conjuring entirely novel pathogens. The AI-designed genomes do, however, carry distinct features that would be difficult to arrive at through natural evolutionary processes alone.

From proteins to whole genomes

The work builds on an approach that mirrors large language models, but substitutes DNA sequences for text. Where a language model predicts the next word or token from a vast corpus of human writing, a large genome model predicts the next nucleotide base - A, T, C or G - from a large collection of genome sequences. Earlier applications showed that prompting such a model with part of a bacterial gene cluster would produce DNA encoding proteins with related functions, including working proteins unlike anything previously catalogued.

Complex-cell genomes proved harder to exploit. Because genes can sit almost anywhere in a eukaryotic genome, outputs that resemble gene-regulatory structures cannot easily be tested for function. As a precaution, the Evo developers deliberately excluded sequences from viruses that target complex cells during training, reasoning that even outputs humans cannot yet interpret could carry unforeseen risk.

Bacterial viruses - bacteriophages - presented a cleaner target. They cannot infect humans, their genomes are compact, and functional tests are straightforward. That made whole-genome generation for a bacteriophage a practical and relatively contained experiment.

The researchers do not stop at the immediate result. They argue that the same class of model, trained differently or prompted differently, could in principle be extended toward viruses that infect vertebrates. Because most of the biological knowledge required to train such a model is already publicly available, the Stanford team suggests that biosecurity thinking should begin now - before that capability exists - rather than after the fact.

// read moreAI Designs 16 Novel Viruses from Scratch in First ExperimentNewsStanford Builds Platform to Track How People Actually Use AINewsGoogle Releases AlphaGenome Atlas to Map Human Genome MutationsNews

Kinetyra