AI can design a protein that has no close natural counterpart, which makes it hard to tell from its sequence alone whether it is meant to help or harm. Google DeepMind’s SynthIDBio aims to add a clue: a digital watermark built into the amino-acid sequence, without changing how tested proteins performed.
A watermark made of amino acids
Existing screening can compare orders for custom DNA synthesis with databases of known biological threats. But a newly designed protein may not resemble anything in those databases. The sequence could point to a useful molecule or to a dangerous one, and familiar matching methods may not know what to make of it.
SynthIDBio adapts Google’s SynthID watermarking approach, previously used for images and text, to protein design. The system works with ProteinMPNN, which builds a protein sequence one amino acid at a time to fit a desired structure.
Keyed choices: At each step, SynthIDBio proposes an amino acid based on a cryptographic key and the choices already made. ProteinMPNN uses the suggestion only when it fits the protein’s structure and is compatible with its function. The result is a sequence selected from among functional possibilities, with watermark-related choices distributed along its length.
Checking for the mark is not a simple yes-or-no match. The verifier estimates whether the sequence contains the amino acids the key would predict more often than chance would suggest. That makes the watermark a statistical signal, not a label that announces itself.
What the tests show, and what they do not
The team tested the method on proteins that bind to several natural targets. In those tests, watermarked versions performed as well as their unmarked counterparts. That is encouraging evidence that the mark need not disrupt function in the examples tested, not a guarantee for every protein a model might produce.
There are practical limits. A very short protein may contain too few watermark-bearing choices for reliable detection. Joining a marked protein to a naturally occurring protein, such as a fluorescent one, can also weaken the signal. And SynthIDBio currently works with ProteinMPNN and similar systems that choose amino acids sequentially. It is not a universal detector for AI-designed proteins.
A screening signal, not a safety certificate
The proposed use is to give companies that synthesize DNA a way to check unfamiliar protein sequences against keys distributed to universities and large biotechnology companies. A match to a trusted key could help explain a sequence’s origin. An apparently AI-designed sequence without a recognized mark could prompt closer review.
That distinction matters. A watermark can help identify a marked design, but its absence does not establish that a sequence is dangerous, and its presence does not by itself prove that a design is safe. The practical gain is narrower: screeners may be able to separate some known, marked designs from sequences that deserve more attention.
SynthIDBio therefore adds a useful layer to biological screening, while leaving the hard judgment calls in place. Even in protein design, a watermark can identify who left a mark. It cannot tell us, on its own, whether the marked protein is a good idea.