Google DeepMind presented SynthID Bio on 30 September 2026, watermarking for synthetic biology. The tool writes a hidden signal into protein sequences and predicted 3D structures without breaking how they work. The methods paper appeared in Nature, and the code, laboratory data and model weights went out to the research community.
Key takeaways
- Launched 30 September 2026 under the name SynthID Bio
- Two data types are marked: amino acid sequences and predicted 3D structures
- For protein folding the watermark is sewn directly into the weights of AlphaFold 3's diffusion network
- In vitro tests covered three targets: VEGF-A, the SARS-CoV-2 spike RBD and PD-L1
- The methods paper ran in Nature, with code and weights released publicly
Two ways to hide a signal
For protein sequences the system gently steers amino acid choice, leaving a recognisable pattern. For predicted spatial structures it shifts atomic coordinates. In the folding case the marking is not a layer on top of the result but part of the model itself, because the team fine-tuned AlphaFold 3's diffusion network so the signal emerges alongside the prediction.
| What is marked | How the mark works | What was verified |
|---|---|---|
| Protein sequence | gentle steering of amino acid choice | hit rates, affinity and diversity match unmarked output |
| Predicted 3D structure | shifted atomic coordinates | prediction accuracy preserved |
| Protein folding (AlphaFold 3) | mark sewn into the diffusion network weights | detectability described as near-perfect |
Does a watermark break the protein
That is a real question, since changing a sequence means changing a molecule. The team compared marked and unmarked binders across three targets and reports comparable hit rates, binding affinity?Binding affinity: A measure of how tightly a molecule holds onto its target. Higher affinity means a more durable bond and usually a stronger effect. and sequence diversity.
For the AlphaFold 3 variant DeepMind describes detectability as near-perfect with prediction accuracy preserved. Adaptyv Bio ran the in vitro validation, and the bacteriophage?Bacteriophage: A virus that infects bacteria only. In synthetic biology it serves as a vehicle for carrying and testing designed sequences. work involved the Hie Lab at Stanford University and the Arc Institute.
From images to molecules
SynthID already covers images, video, audio and text, including the Gemini app, Lyria and NotebookLM. The SynthID Detector portal is still in testing with journalists. Extending it to biology carries the same idea into material no one can check by eye. Sarah Carter of Science Policy Consulting, who reviewed the project early, calls it "an important piece of the puzzle for tracking the provenance of biological designs".
Why it matters
Tracing where a sequence came from is harder than tracing a photograph, and the stakes are higher. DNA synthesis companies screen orders for hazards, but they cannot see whether a design came out of a model. A watermark built into the weights gives that screening a handle. The weak link is that it is voluntary, because only those looking for the signal will find it, and only in designs from marked models.
What's next?
- Code, in vitro data and model weights are already available, letting independent teams test how well the mark survives sequence edits
- SynthID Detector stays in closed testing, and Google has given no date for general availability
Sources
- Google DeepMind — Introducing SynthID Bio
- Google DeepMind — SynthID
