DeepMind's SynthID Bio marks AI proteins without losing function

Google DeepMind had already developed SynthID to embed hard-to-perceive marks in images, audio, video and text. SynthID Bio extends that idea to synthetic biology as a proof of concept. The body frames the need directly: as designing new proteins becomes easier, it becomes harder to tell whether an unfamiliar sequence came from nature or from an AI. DNA-synthesis companies compare ordered sequences against databases of known dangerous sequences, but AI can generate sequences that barely resemble known ones, so matching alone may struggle to establish origin. Unlabeled AI-generated 3D structures entering public databases could also affect follow-on research.

The technical approach is deliberately light-touch. For amino acid sequences, the system slightly biases amino acid choices during generation within a range that does not greatly change protein properties, leaving a statistically detectable feature. For structural data, atomic coordinates are subtly adjusted to carry similar identifying information. In tests combining AlphaProteo and ProteinMPNN against three targets — VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1 — the watermarked proteins showed design success rates and binding strength similar to unwatermarked ones, with sequence diversity maintained. DeepMind also fine-tuned part of AlphaFold 3's diffusion network so the model itself generates marks, and reported almost complete detection while maintaining prediction accuracy, even with small noise or coordinate changes.

DeepMind positions SynthID Bio as useful for DNA-synthesis screening and for identifying AI-generated data in databases such as Protein Data Bank, UniProt and GenBank. Along with Stanford University and the Arc Institute, it has also embedded marks in genomes designed by the genome-generating AI Evo 2, and early experiments confirmed watermarked bacteriophages functioned in bacterial culture. Whether this approach becomes routine likely hinges on how well the marks survive deliberate modification and how they are paired with provenance metadata and registration systems; a technical paper on extending the method to more complex biological data is planned.