DeepMind watermark can flag proteins made by AI

Google DeepMind researchers have developed SynthIDBio, a watermarking system that embeds a hidden statistical signature into proteins designed by AI systems such as AlphaFold and RFdiffusion. The marker is encoded in amino-acid sequences and in tiny changes to protein shape, while tests found watermarked proteins could still bind targets linked to viral infection, blood-vessel formation and immune regulation as efficiently as unwatermarked versions.

The system is intended to help distinguish AI-generated proteins from natural ones in biological repositories and to give DNA synthesis providers another signal when screening customer orders. A secret detection key would be shared with trusted partners, allowing them to flag proteins made by AI even when a sequence has little resemblance to known toxins, pathogen proteins or other molecules in reference databases.

Researchers caution that the mark is not a complete biosecurity solution. A watermarked protein can often be passed through another design tool to produce a new sequence that preserves structure and function while obscuring its origin, making the technique one layer in a broader safety framework.