Google DeepMind introduces SynthID Bio, watermarking AI-designed proteins without hurting function in wet-lab tests, with code and weights released to researchers

Why it matters

【Key points】

On Sep 30 Google DeepMind introduced SynthID Bio, bringing SynthID watermarking to synthetic biology. It is a proof of concept: an imperceptible signature in AI-designed protein sequences and predicted 3D structures, verifiable on the synthesized protein, without harming biological function. DeepMind is publishing a methods paper, open-sourcing code and in vitro data, and releasing weights to researchers.

【Verifiable】

For sequences, SynthID Bio subtly guides amino-acid choices; for structures, it fine-tunes a small part of AlphaFold 3's diffusion network so predicted coordinates carry a detectable signature, with accuracy preserved and near-perfect detectability reported. In wet-lab tests, binders designed with AlphaProteo and a watermarked ProteinMPNN against VEGF-A, the SARS-CoV-2 spike RBD and PD-L1 matched unwatermarked designs on hit rate, binding affinity and sequence diversity. With the Hie lab at Stanford and Arc Institute, DeepMind also watermarked the genome of an Evo 2-designed bacteriophage; early tests show the phages remain functional, with a manuscript to follow. DeepMind says robustness to deliberate tampering is the next challenge.

【Our read】

A watermark can show that a design came from a trusted model with safeguards; it cannot catch a malicious design made with an unwatermarked open model or modified weights. That makes it a fast lane for legitimate DNA synthesis orders rather than a net for bad ones. Its impact depends on synthesis providers building the signal into screening and on other protein-design developers adopting it. Releasing code and weights signals DeepMind wants an industry standard, not a Google-only feature.