Google DeepMind published a method in Nature on 30 September for watermarking proteins designed by AI. SynthID Bio is a family of techniques that weave a faint statistical signal into both the amino acid sequence and the predicted three-dimensional structure of a designed protein, without noticeably compromising its stability, function or the speed of generating it.

The problem it addresses is provenance. Protein design models are now good enough that a sequence arriving at a DNA synthesis provider carries no reliable indication of whether a person or a model wrote it. Screening providers check ordered sequences against databases of known hazards, which works for things already catalogued and not for novel designs.

What the method does

Rather than appending a tag, the watermark steers generation: the model is nudged toward amino acid choices that carry a detectable statistical pattern while still landing on a protein that folds and works. DeepMind reports the designed proteins retain structural stability and functional efficacy, and that generation speed is not meaningfully affected.

An abstract molecular model made of coloured spheres
The mark is woven into sequence and structure. Illustrative photo. Google DeepMind · pexels · Pexels License

The company is pairing it with practice rather than leaving it as a paper. It has a partnership with Twist Bioscience to fold watermark detection into DNA synthesis screening, which is where a mark of this kind is actually useful — at the point an order is placed.

Gloved hands holding a small vial
DNA synthesis screening is the intended point of use. Illustrative photo. Thirdman · pexels · Pexels License

The limitation is stated plainly

The watermark can be scrubbed. Nature’s accompanying news piece notes that someone who wants to remove the mark can in many cases run a watermarked protein through another design tool and obtain a new sequence that keeps the structure and function while losing the signal. That makes this a provenance tool for the cooperative case, not a control on a determined actor.

That is still worth having. Most sequences arriving at a synthesis provider are not sent by a determined actor, and a screen that can say where a design came from narrows the set that needs a human to look at it.

What to watch

Whether other protein design labs adopt a compatible marking scheme, and whether any biosecurity regulator starts referring to watermarking in synthesis screening guidance rather than leaving it to providers.