Researchers at Stanford University have used large genome models — AI systems trained on DNA sequences rather than text — to generate functional viral genomes. The AI-designed viruses are closely related to existing bacteriophages but exhibit distinct genetic features that would be difficult to achieve through natural evolution. The work raises biosecurity concerns, with the researchers suggesting the field should proactively consider the risk of similar AI systems being used to design viruses targeting vertebrates.
Questions this post answers
What are large genome models and how do they differ from large language models?
Large genome models apply the same next-token prediction training approach as large language models, but use DNA sequences instead of text. DNA uses only four letters (A, T, C, G), which simplifies the alphabet, but the biological context of each position varies enormously — some positions are highly constrained while others are nearly arbitrary — making the task non-trivial. Fed enough genome sequences, these models learn to output biologically coherent DNA. Researchers tracking AI applications in genomics and biosecurity follow developments like these on daily.dev.