Google DeepMind has extended its SynthID watermarking family into synthetic biology, publishing a method that embeds an invisible, verifiable signature inside AI-designed proteins.
SynthID Bio targets two things at once: the amino acid sequence a model writes and the 3D structure it predicts. The goal is to let researchers trace a design back to the system that produced it without changing what the protein does.
The proof of concept held up in wet-lab work on three target proteins. Watermarked designs performed comparably to unwatermarked versions and kept the diversity found in natural sequences, the company said.
That matters because protein design tools have become cheap and powerful. DeepMind points to AI-designed enzymes that digest plastic and proteins that block venom, then notes the same capability could be misused.
The team frames watermarking as a biosecurity layer rather than a lock. A detectable mark gives labs, publishers and journals a way to check provenance, and it settles credit disputes because a design carries its own evidence of origin.
The work was published in Nature. DeepMind calls it a first step, and says any watermark is only useful if the wider field adopts and tests it.
