Google DeepMind Unveils SynthID Bio to Watermark AI-Designed Proteins Without Affecting Function

As generative AI tools such as AlphaFold and AlphaProteo enable the design of novel proteins and bacteriophages, they also introduce risks. AI-designed sequences may bypass traditional DNA synthesis screening, and mislabeled 3D structures can pollute public databases. SynthID Bio aims to provide an automated verification layer to mitigate these risks by allowing DNA synthesis providers to more efficiently screen for potential threats.

The method adapts based on the data type: it subtly guides amino acid selection for sequences and adjusts atomic coordinates for predicted 3D structures. In wet-lab testing involving target proteins such as the SARS-CoV-2 spike protein RBD and VEGF-A, the watermarked designs maintained the same hit rate, binding affinity, and natural sequence diversity as unwatermarked versions.

For protein folding tasks, SynthID Bio fine-tunes parts of the AlphaFold 3 diffusion network, embedding watermarks directly into the model's weights. This ensures that predicted 3D coordinates inherently carry a detectable signature regardless of the user.