DeepMind Unveils AI-Powered Watermark for Designed Proteins

Google's DeepMind has developed SynthID Bio, an AI-powered watermark for designed proteins that preserves their functionality. This breakthrough is significant in the field of biosecurity as it enables researchers to verify the origin of unfamiliar protein sequences without manually checking each order.

The SynthID Bio method conceals a signature within the amino acid sequence and 3D structure of the protein, which can be detected after synthesis. In lab tests, DeepMind paired its AlphaProteo design method with SynthID Bio and ran experiments on three targets: VEGF-A, SARS-CoV-2 spike protein receptor-binding domain, and PD-L1. The results showed that watermarked designs matched unwatermarked versions in terms of hit rate, binding affinity, and natural sequence diversity.

DeepMind also integrated SynthID Bio into its AlphaFold 3 structure prediction model to embed the watermark in predicted coordinates. This approach maintains the model's accuracy while allowing for near-perfect detectability of the signature. The team suggests that this method could be used to flag synthetic entries in public databases, such as the Protein Data Bank and UniProt.

However, DeepMind acknowledges that resistance to deliberate tampering is still a challenge to be addressed. They propose combining SynthID Bio with provenance metadata or central repositories of AI-generated biological data to strengthen biosecurity measures.