Google DeepMind says its SynthID Bio watermark can be embedded in AI-designed protein binders without reducing their measured binding performance. In lab tests across three targets, the team reports marked designs bound about as often and as strongly as unmarked ones. The result is an early provenance technique, not a biological-safety check.
The sequence watermark works by subtly steering amino-acid choices as a protein is generated, leaving a pattern that a detector can recognize. DeepMind combined its AlphaProteo binder-design method with a SynthID Bio-enabled version of ProteinMPNN, a model that generates protein sequences.
Wet-lab tests covered designs aimed at VEGF-A, the SARS-CoV-2 spike protein's receptor-binding domain, and PD-L1. DeepMind reports similar hit rates and binding affinities for marked and unmarked designs, along with similar sequence diversity; its announcement does not define that diversity measure in the summary.
A separate method marks predicted protein structures rather than binder sequences. The team fine-tuned a small part of AlphaFold 3's diffusion network so its predicted atomic coordinates carry a detectable signal. DeepMind says this approach preserved prediction accuracy and remained detectable under digital noise and minor coordinate changes; those structure-prediction tests are distinct from the wet-lab binder results.
The watermark is intended to help identify where a biological design came from. DeepMind says resistance to deliberate tampering remains an open challenge. A detected mark can support a provenance claim, but by itself it does not determine whether a protein is safe or harmful.
DeepMind says it is publishing the methods paper, open-sourcing code and in-vitro data, and releasing model weights to researchers.
Sources
- Nature paper: “Function-preserving watermarking of AI-generated proteins”
- Google DeepMind's SynthID Bio announcement and methods summary
Image: PD-L1 protein structure (PDB 3BIK), by Goultard59, via Wikimedia Commons; CC BY-SA 4.0 (license). The illustration shows the target protein, not a watermarked binder or an image from the study.
