Google DeepMind introduces invisible watermarks for AI-designed proteins

Google DeepMind has introduced SynthID Bio, a tool designed to embed invisible cryptographic signatures directly into proteins designed by artificial intelligence. This technology aims to enhance biosecurity and improve the traceability of molecular designs in an era where AI can generate amino acid sequences that do not exist in nature.

Traditional security systems for DNA synthesis rely on comparing requested sequences against databases of known threats. However, these systems may fail to identify entirely new, AI-generated sequences. SynthID Bio addresses this by incorporating a watermark into the molecular design itself. For protein sequences, the system works within ProteinMPNN, the industry-standard design tool, by suggesting specific amino acids that form a statistical signal detectable by software without compromising the protein's biological function.

For three-dimensional structures, the system adjusts atomic coordinates within the AlphaFold 3 diffusion network, embedding the signature into the model's weights. The project includes an open-source release for researchers to facilitate widespread adoption and scientific oversight.