Google DeepMind has introduced SynthID Bio, a family of methods for marking AI-generated protein sequences and predicted three-dimensional structures. The goal is to make those designs traceable to their origin.
The team’s Nature paper, published Sept. 30, reports functional watermarked protein binders with binding affinity comparable to unwatermarked counterparts. The authors describe the work as a proof of concept.
A signature in the biological design
DeepMind explains that its approach subtly guides the choice of amino acids in protein sequences and adjusts atomic coordinates in predicted structures to create a detectable signal. The sequence method works with ProteinMPNN; the structure method uses a fine-tuned AlphaFold 3 model, according to the paper.
The laboratory tests examined binders, molecules designed to attach to other proteins. The researchers tested them against three targets: VEGF-A, PD-L1 and the SARS-CoV-2 spike protein’s receptor-binding domain. Their results support preserving binding function in these tested designs.
Practical uses still need work
DeepMind proposes using these signals to help DNA synthesis providers screen orders and to flag synthetic entries in scientific databases. Those are potential uses, rather than established deployments: the paper says putting them into practice will require further innovation, coordination and standardization across the industry.
The announcement also identifies resistance to deliberate tampering as an ongoing challenge. DeepMind presents watermarking as one layer of biosecurity, alongside other safeguards.
