← All news

Latest AI news · Original source attached

SynthID Bio watermarks proteins, not just files

Google DeepMind's SynthID Bio embeds a verifiable signature in biological code, with wet-lab validated protein binders and AlphaFold 3 weight watermarking.

Reported from Google DeepMind

Summary written from the original first-party source — the link to it is at the end of this post.

Google DeepMind extended SynthID into synthetic biology with SynthID Bio, a family of watermarking methods that embed an imperceptible signature into biological code, verifiable on the synthesized physical protein rather than only a digital file. That distinction is the whole point: a text watermark dies when the text is paraphrased, a protein watermark has to survive translation into matter.

Two mechanisms are involved. One subtly steers amino-acid choice in sequences. The other adjusts atomic coordinates in predicted 3D structures.

For sequences, the work was validated through AlphaProteo plus a SynthID-enabled ProteinMPNN and wet-lab tested on three targets: VEGF-A, SARS-CoV-2 spike RBD and PD-L1. Watermarked designs matched unwatermarked versions on hit rate, binding affinity and sequence diversity, which Google calls the first watermarked and biologically functional protein binders.

For structures, the method fine-tunes part of AlphaFold 3’s diffusion network so the signature is baked into the weights, meaning the coordinate signal survives whoever runs the model. Google reports preserving AlphaFold 3’s prediction accuracy with near-perfect detectability, holding up against digital noise and minor coordinate changes, with 7PPA as the worked example. DeepMind also extended the work with the Hie lab at Stanford and Arc Institute into Evo 2, watermarking the genome of an Evo 2-designed bacteriophage and confirming it functional in bacterial culture.

Google calls this a proof of concept and a first step, and names the hard remaining problem directly: robustness against deliberate tampering. The claims about DNA-synthesis screening and flagging synthetic entries in PDB, UniProt and GenBank are prospective rather than deployed. The bacteriophage details are promised in a technical manuscript that is not published yet. The methods paper is in Nature, and code plus in vitro data are open sourced with weights to be released to the research community.

(Original source: Google DeepMind)

AI Slop Weekly

Get this story in context, every Sunday.

The monthly rundown of what the stories above add up to — what shipped, what changed, what to actually use. 3-minute read, free forever.

Subscribe free

No spam. Unsubscribe anytime.