Google DeepMind's SynthID Bio puts a hidden watermark into AI-designed proteins

Google DeepMind has developed SynthID Bio, a method that hides a watermark inside AI-designed proteins, SingularityHub reported on 5 October 2026.

Why it matters

AI tools can now design entirely new proteins — potential drugs, but also, in the wrong hands, possible toxins. DNA synthesis companies screen orders against databases of dangerous sequences, but AI-designed sequences can look unlike anything known and slip through. Researchers also worry about unlabelled AI-generated structures polluting scientific databases.

How it works

SynthID Bio is adapted from SynthID-Text (2024), which nudges a model’s word choices to leave a statistical signature. For proteins, it subtly steers design decisions such as which amino acid comes next and makes tiny shifts in atom positions. It fine-tunes a small part of AlphaFold3, so proteins from the model carry the mark regardless of who runs it; it could be built into tools such as AlphaFold and RFdiffusion.

Results and limits

• In lab tests, watermarked protein binders aimed at targets linked to immune regulation, blood-vessel growth and viral proteins performed as well as unmarked versions — the team calls them “the first-ever watermarked and biologically functional protein binders”.
• The mark can be scrubbed: running a protein through another design tool to get a similar sequence effectively removes it.
• It cannot identify who made a protein, limiting its use for tracing misuse or marking intellectual property.
• Oxford researcher Oliver Crook told Science that blanket watermarking could add complexity for non-sensitive research.
• Who should hold the detector key is unresolved; a survey of 130 stakeholders reached no consensus.

DeepMind is releasing the code and experimental data to researchers and, with Arc Institute and Stanford, has in early tests watermarked the genomes of AI-designed bacteriophages without disabling them.