Google DeepMind (opens in new tab) introduced SynthID Bio on Sept. 30, a family of watermarking methods that embed a detectable signature in AI-generated protein sequences and predicted 3D structures while aiming to preserve the proteins’ biological function. The method is described in a paper published in Nature (opens in new tab), according to Nature (opens in new tab) and Science (opens in new tab).
According to DeepMind’s research blog, the system subtly guides the choice of amino acids for sequences and adjusts atomic coordinates for predicted structures, and the watermark can be verified on the synthesized physical protein as well as on digital designs. In wet-lab tests using AlphaProteo and a SynthID Bio–enabled version of ProteinMPNN, watermarked binders against VEGF-A, the SARS-CoV-2 spike protein’s receptor-binding domain and PD-L1 matched the hit rate, binding affinity and natural sequence diversity of unwatermarked versions, DeepMind said.
For structure prediction, DeepMind said it fine-tuned a small part of AlphaFold 3’s diffusion network so that predicted coordinates carry a signature, preserving prediction accuracy while offering near-perfect detectability. DeepMind says the watermark could give DNA synthesis providers an automated signal that an order came from a trusted model, and could help flag synthetic entries submitted to public databases such as the Protein Data Bank, UniProt and GenBank. Nature (opens in new tab) reported that the mark could be detected by anyone holding a secret key, which would be shared only with trusted partners such as DNA sequence makers.
DeepMind says it is publishing its methods paper, open-sourcing the code and in vitro data, and releasing the weights, and it lists making the watermark more robust against deliberate tampering as a key challenge. Nature (opens in new tab) reported that in many cases someone could erase the tag by running a watermarked protein through another design tool, and that biosecurity researcher Tessa Alexanian sees the method as one more tool in a layered defense against biological threats. DeepMind also said that in ongoing work with the Hie lab at Stanford University and Arc Institute, it watermarked the genome of a bacteriophage designed with the Evo 2 model, and that early lab tests showed the watermarked phages were functional.