Google DeepMind published a methods paper in Nature on 30 September describing SynthIDBio, a family of watermarking techniques for AI-generated proteins. All twenty-one listed authors are at Google DeepMind in London. There are two variants: SynthIDBio-sequence embeds a watermark into the amino acid sequence itself, and SynthIDBio-structure is a fine-tuned AlphaFold 3 model that embeds an imperceptible watermark into predicted biomolecular structures.
The difficulty is that in a protein the sequence is the function, so any change risks destroying what the protein was designed to do. The paper reports watermarked designed binders with “binding affinity comparable with non-watermarked counterparts” and “near-perfect watermark detection accuracy”. DeepMind’s accompanying post names the three targets tested as VEGF-A, the SARS-CoV-2 spike protein receptor binding domain and PD-L1.
Why it matters: provenance. A synthesis provider that can test whether an ordered sequence came from a model with safeguards built in gains a screening signal it does not currently have, which is a biosecurity argument rather than a scientific-integrity one.
The authors describe the work as a proof of concept, and that framing is doing real work. Detection accuracy on three targets in a laboratory is not robustness against someone deliberately stripping a watermark, and the paper does not claim otherwise.
Source: the paper in Nature.
