Google DeepMind watermarks AI-designed proteins

Google DeepMind has released SynthID Bio, a set of watermarking methods for synthetic biology. It brings the company's SynthID provenance tool to AI-generated biological sequences and structures. DeepMind is publishing a methods paper, open-sourcing the code and in vitro data, and releasing the weights to researchers.

DeepMind describes the watermark as one layer in a "Swiss cheese" biosecurity model, alongside model-level mitigations and customer vetting. It is aimed mainly at DNA synthesis screening. Screeners used to assume that an unfamiliar sequence was probably an undiscovered natural organism. That assumption no longer holds when AI can generate novel sequences that look nothing like known hazards, and such orders can now require slow manual review. According to DeepMind, a watermark could give screeners an automated signal that an order came from a trusted model with built-in safeguards.

James Diggans, VP of Policy and Biosecurity at Twist Bioscience, gave early feedback on the paper. He called watermarking "a promising new addition to the biosecurity toolbox." DeepMind also suggests the method could help public databases such as the Protein Data Bank, UniProt and GenBank label or flag synthetic entries.

DeepMind acknowledges that the watermark is not yet robust against deliberate tampering. It says the approach could be combined with C2PA-style provenance metadata or with central repositories of AI-generated biological data.

Separately, DeepMind worked with the Hie lab at Stanford and Arc Institute to integrate SynthID Bio into the Evo 2 genomic model and watermark a bacteriophage genome designed by Evo 2. Early tests in bacterial cultures confirmed that the watermarked phages are functional. DeepMind has not released those results yet and says a technical manuscript will follow.