Google DeepMind has developed a way to watermark proteins designed by artificial intelligence without disrupting their biological function in laboratory tests.
The research, published in Nature on September 30, introduces SynthID Bio, a family of methods for marking protein sequences and predicted structures with detectable signatures.
For sequences, the system subtly guides the choice of amino acids, the building blocks of proteins. For predicted structures, it adjusts atomic coordinates to create a signal that detection tools can recognize.
The team tested protein binders, molecules designed to attach to particular protein targets, using AlphaProteo alongside a modified version of ProteinMPNN. The experiments showed that a watermark could survive the move from a digital design to a physical protein while retaining biological activity.
That matters as AI increasingly contributes designs to scientific databases and biotechnology research. Knowing how a protein was produced could help researchers track its origins and support checks by DNA synthesis providers.
The work remains a proof of concept. A detectable signature identifies provenance; it does not establish that a protein is safe. Its immediate contribution is demonstrating that traceability and useful biological function can coexist.