Introducing SynthID Bio

Generative AI is increasingly used to tackle biological problems, from predicting protein structures with AlphaFold to designing entirely new proteins with AlphaProteo and ProteinMPNN, and more recently to developing new bacteriophages that infect bacteria. But these tools create new risks: novel AI designs can bypass traditional DNA synthesis screening, and mislabeled synthetic 3D structures risk polluting public databases and misleading downstream research. SynthID Bio is DeepMind's answer, extending watermarking into synthetic biology by embedding an imperceptible signature directly into the biological code, so the watermark can be verified not just in a digital model but on the synthesized physical protein itself.

SynthID Bio is a family of watermarking methods built specifically for synthetic biology to strengthen biosecurity and scientific integrity. It adapts its approach depending on the data type, subtly guiding the choice of amino acids for protein sequences and adjusting atomic coordinates for predicted 3D structures, thereby creating a reliable signal for detection. Crucially, these adjustments did not compromise the protein's biological function in laboratory testing — an essential requirement if watermarked designs are to remain useful for treating disease and advancing research.

The team verified the approach on protein binders, molecules built to selectively latch onto other proteins, using the binder design method AlphaProteo together with a SynthID Bio-enabled version of ProteinMPNN, the commonly used protein sequence generation method. In wet-lab testing across three target proteins — VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1 — the watermarked designs matched the hit rate, binding affinity (measured as KD, where lower means stronger binding), and natural sequence diversity of unwatermarked versions, producing the first-ever watermarked and biologically functional protein binders.

For protein folding, SynthID Bio fine-tunes a small part of AlphaFold 3's diffusion network, building the watermarking ability directly into the model's weights so that predicted 3D coordinates inherently carry a detectable signature regardless of who runs the model. This preserves AlphaFold 3's prediction accuracy while offering near-perfect detectability, maintaining key structural feature distributions and holding up against digital noise or minor coordinate changes.