Google DeepMind Introducing SynthID Bio

Google DeepMind has introduced SynthID Bio, a watermarking system designed to identify AI-generated biological sequences and 3D structures. This technology allows for the verification of synthetic proteins not only in digital models but also in their physical, synthesized forms, providing a critical layer of provenance for synthetic biology.

Technical Implementation of SynthID Bio

SynthID Bio employs different watermarking methods depending on the type of biological data being processed to ensure a detectable signal is created without altering the protein's biological function.

Protein Sequence Watermarking

For protein sequences, the system subtly guides the selection of amino acids. This process was verified using a SynthID Bio-enabled version of ProteinMPNN (a common protein sequence generation method) in conjunction with AlphaProteo for designing protein binders.

3D Structure Watermarking

For protein folding and 3D structures, SynthID Bio fine-tunes a small portion of AlphaFold 3’s diffusion network. By embedding the watermark directly into the model's weights, the predicted 3D coordinates inherently carry a detectable signature. This method maintains AlphaFold 3's prediction accuracy and remains robust against digital noise or minor coordinate adjustments.

Experimental Validation and Performance

Wet-lab testing was conducted on protein binders designed to latch onto three target proteins: VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1. The results demonstrated that watermarked designs matched unwatermarked versions in three key metrics:

• Hit rate: The frequency of successful binders.
• Binding affinity: Measured as $K_D$, where lower values indicate stronger binders.
• Natural sequence diversity: The variety of sequences generated.

Biosecurity and Information Integrity Implications

SynthID Bio is designed as part of a "Swiss cheese" defense model, adding a tangible verification layer to biological designs to address gaps in other safety measures.

DNA Synthesis Screening

DNA synthesis providers screen requests against databases of known threats. However, because generative AI can create entirely new sequences that do not resemble known hazards, manual reviews can be time-consuming. SynthID Bio provides an automated verification signal that can prove a sequence originated from a trusted model with built-in safeguards, streamlining the screening process.

Database Integrity

The technology aims to prevent the pollution of public biological databases (such as UniProt, GenBank, and the Protein Data Bank) by ensuring that AI-generated entries are properly labeled or flagged, reducing the risk of mislabeled synthetic structures misleading downstream research.

Future Directions and Research

Google DeepMind is expanding the application of SynthID Bio to more complex biological objects. In collaboration with the Hie lab at Stanford University and Arc Institute, the team integrated SynthID Bio into Evo 2, a genomic model, to watermark the genome of a designed bacteriophage. Early laboratory testing has confirmed that these watermarked bacteriophages remain functional.

To further the research, Google DeepMind is open-sourcing the code, releasing the weights, and publishing the methods paper and in vitro data to the research community.