AI Learns to Write Viral Genomes

Researchers set two AI models a task that looks both promising and unsettling: to generate the complete genome of a viable bacteriophage — a virus able to penetrate a bacterium, multiply inside it and destroy it from within.

As a template they took the already known bacteriophage ΦX174 (pronounced “phi-X-one-seven-four”), which can infect and kill the bacterium E. coli. Building on this example, the models proposed around 700,000 possible genome variants. Out of this whole set the researchers selected 285 designs that looked the most promising.

From there the work moved from the digital world into the laboratory. Based on the selected designs, new DNA molecules were synthesized and introduced into E. coli cells, after which all that remained was to wait: would viable bacteriophages emerge from these artificially assembled sequences?

The very fact that generative models can not merely analyze biological sequences but also propose workable constructions of entire viral genomes shows just how far machine-learning tools have advanced in their application to molecular biology. This is precisely why research of this kind evokes a dual feeling: admiration for what it can do and wariness about where it opens the door.

Source: Schneier on Security