Google's AI Lab Hides a Traceable Signature Inside AI-Designed Proteins, Without Breaking Them

AI programs can now design brand-new proteins, the tiny working parts of living things, and some worry that a harmful design could slip past the safety checks meant to catch it. Google DeepMind says it can now hide an invisible signature inside those designs that shows they came from an AI, and its lab tests found the marked proteins still work just as well.

Google DeepMind, Google's AI research lab, announced a tool today called SynthID Bio. It puts a hidden signature, a kind of watermark, into proteins designed by AI.

Proteins are the tiny machines that do most of the work in living things. Each one is a chain of small building blocks, and the recipe for it is written in DNA. AI programs can now design proteins that have never existed in nature, which could lead to new medicines. But it also raises a worry. Companies that make DNA to order check each request against a list of known dangerous recipes. A brand-new AI design might not look like anything on that list, harmful or not. There's a second worry: AI-made designs could end up in the big public libraries scientists rely on, mislabeled as natural.

SynthID Bio tackles both by leaving a mark. While the AI designs a protein, it nudges some of the choices about which building blocks go where, in a pattern too subtle to notice but that a checking program can pick up. DeepMind says the mark can be read not only in the digital design but in a real protein made from it.

The big question was whether the mark would spoil the protein. In lab tests, DeepMind's team designed proteins meant to latch onto three different targets, one of them part of the virus that causes COVID-19. The marked versions grabbed their targets just as well as unmarked ones. Working with Stanford University and the Arc Institute, a research group, the team also marked the DNA of viruses that infect bacteria, designed by another AI program, and early tests found those viruses still worked. DeepMind is publishing the method in the journal Nature and sharing its code and data with other researchers.

DeepMind says plainly that this is not a complete answer. Making the mark hard to remove on purpose is still an open problem, and a mark can only appear in designs made by programs that build it in. It calls the watermark one layer of protection among several. Outside the company, James Diggans of Twist Bioscience, a firm that makes DNA to order, called it "a promising new addition to the biosecurity toolbox," and biosecurity policy expert Sarah Carter called it "an important piece of the puzzle."