Google DeepMind Launches SynthID Bio for Protein Watermarking

Google DeepMind has introduced SynthID Bio, a digital watermarking method tailored for synthetic biology. Unlike image or text watermarks, this signal extends beyond digital files and embeds directly into proteins synthesized as physical molecules. Laboratory tests show that watermarked proteins retain their original biological functions without disruption.

The method inserts subtle signals by adjusting data types. For protein sequences, it guides amino acid selection; for 3D structures, it shifts atomic coordinates slightly. These modifications leave enough trace for later detection while preserving functionality. The research team tested this concept using protein binders designed via AlphaProteo and ProteinMPNN.

Stock photo for illustration only, not from the actual event

In vitro experiments across three target proteins—VEGF-A, the SARS-CoV-2 receptor binding domain, and PD-L1—demonstrated that watermarked proteins matched original benchmarks in design success rate, binding affinity, and sequence diversity. These are the first protein binders that combine watermarking with viable biological activity.

Additionally, the team tuned AlphaFold 3 network weights so predicted 3D coordinates automatically carry watermarks. Model accuracy remains unaffected, and watermarks resist digital noise. This approach reinforces biosecurity using a Swiss Cheese Model framework, layering multiple safeguards to close security gaps.