Google DeepMind introduces SynthID Bio to watermark AI-designed proteins

Google DeepMind has introduced SynthID Bio, a watermarking technology designed to embed invisible signatures directly into AI-generated protein sequences and structures. This development aims to enhance biosecurity and scientific integrity in the field of synthetic biology.

By subtly adjusting the probability distribution of amino acids during the design process, the technology creates a detectable signal that survives file format changes and distribution. Researchers at the University of Maryland and Google have demonstrated that these watermarks do not compromise the biological function, structural soundness, or folding quality of the proteins. In laboratory tests involving targets such as the SARS-CoV-2 spike protein, watermarked designs performed consistently with non-watermarked versions in terms of binding affinity and diversity.

The system also addresses the risks posed by generative AI, which can create novel biological sequences that might bypass traditional DNA synthesis screening. SynthID Bio provides a layer of verification to identify sequences originating from trusted models. Additionally, the technology can be integrated into models like AlphaFold 3 to embed watermarks into predicted 3D coordinates, helping to maintain the integrity of public protein databases.