SynthID Bio: DeepMind ‘tags’ AI-designed proteins without impairing their function

An invisible signature embedded in the sequence and 3D structure of the proteins generated is intended to help DNA manufacturers identify orders originating from models equipped with safeguards.

Google DeepMind has published in Nature SynthID Bio, a family of AI-designed watermarking methods for proteins. The mark is inscribed on two levels: in the amino acid sequence, by discreetly directing the choice of the design model (a modified version of ProteinMPNN) at each position, and in the atomic coordinates of the predicted structures, thanks to a part of the AlphaFold 3 diffusion network retrained for this purpose. The signal can be found on the physical protein once synthesized.

Laboratory tested on three targets

Combined with AlphaProteo, the method has been benchtop tested on three targets: VEGF-A, the receptor binding domain of the SARS-CoV-2 spike protein, and PD-L1. In all three cases, the tattooed binders match the unlabeled versions in success rate, binding affinity, and sequence diversity. The tag is resistant to digital noise and small changes in coordinates.

The goal is biosafety: DNA synthesis providers, who confront each command with databases of known threats, would have an automatic clue that a sequence comes from a trusted model. DeepMind acknowledges that resistance to deliberate tampering remains an open problem and presents the tool as one layer among many. The code, in vitro data, and weights are open to researchers; Stanford's Hie lab and the Arc Institute are already collaborating on applications to bacteriophages.