On September 30, Google DeepMind released SynthID Bio, bringing its existing SynthID watermark to synthetic biology. The company says that once the mark is in a biological design, it can be checked not only on the digital model but on the synthesized physical protein, and that biological function still held in laboratory testing. The post presents this as a proof of concept.
[1]
The wet lab used AlphaProteo and a SynthID Bio-enabled version of ProteinMPNN to design protein binders. The targets were VEGF-A, the receptor-binding domain of the SARS-CoV-2 spike protein, and PD-L1. The company says the watermarked designs matched unwatermarked ones on hit rate, binding affinity (measured as KD, where lower means a stronger binder), and natural sequence diversity. That comparison is DeepMind’s own wet lab, not an independent replication.
On structure prediction, the company says it fine-tuned a small part of AlphaFold 3’s diffusion network so the watermark sits in the model weights. Predicted 3D coordinates then carry a detectable signature no matter who runs the model. The post says prediction accuracy is preserved, detectability is near perfect, key structural distributions hold, and the mark stands up to digital noise or small coordinate changes.
[1]The company describes the watermark as one more verification layer in biosecurity, not a measure that is sufficient on its own. Turning a digital protein design into a physical molecule means ordering from a DNA synthesis provider, which screens requests against databases of known threats. DeepMind writes that AI can create entirely new sequences with little resemblance to known hazards, so a screener can no longer assume an unfamiliar sequence is simply a natural organism that has not been catalogued. SynthID Bio can supply an automated signal that an order came from a trusted model with built-in safeguards. The post also says that, for public-submission databases such as the Protein Data Bank, UniProt, and GenBank, it could help label synthetic entries or flag them for further review.
James Diggans, vice president of policy and biosecurity at Twist Bioscience, gave early feedback. He said that for Twist, watermarking could be one more tool in screening, focusing effort on sequences that need a closer look. That is his judgment, not a screening rule already in place.
[1]The company says no single biosecurity measure covers the whole risk. A key open problem is making the watermark hold up against deliberate tampering. The mark can also be paired with provenance metadata, similar to C2PA for digital media, or with central repositories of AI-generated biological data. Ongoing work with the Hie lab at Stanford and the Arc Institute connects SynthID Bio to the genomic model Evo 2, to mark a bacteriophage that Evo 2 designed. Early tests in bacterial cultures found that these watermarked bacteriophages still function. The details are left to a later technical manuscript.
The company says it is publishing a methods paper and releasing the code, the in vitro data, and the weights. The contact address for collaboration is synthidbio@google.com. The post asks for a high-level proposal only, with no confidential or proprietary information attached.
[1]要点
- SynthID Bio watermarks AI-designed proteins. The company says the mark can be checked on the physical protein and that lab function held.
- Wet-lab targets were VEGF-A, the SARS-CoV-2 spike RBD, and PD-L1. Hit rate, KD, and sequence diversity matched the unwatermarked designs.
- A small AlphaFold 3 fine-tune is said to keep prediction accuracy with near-perfect detectability.
- Resistance to deliberate tampering is unsolved. The company says code, in vitro data, and weights will be released. Bacteriophage details are not in this post.