DeepMind unveils SynthID Bio to watermark AI-designed proteins for provenance and biosecurity

DeepMind said Wednesday it has developed a proof-of-concept system for watermarking AI-designed proteins so they can be identified later without losing their biological function in lab tests, a step the Google-owned company says could help track the origin of synthetic biology and strengthen biosecurity screening. In a Sept. 30 blog post titled “Introducing SynthID Bio,” DeepMind described a family of methods that embeds an “imperceptible signature” into biological code and said that watermark can be verified both in digital models and in a physically synthesized protein. The claims come from a company announcement, not an independently verified peer-reviewed study.

The technical work spans both protein design and structure prediction, according to DeepMind. For protein sequence design, the company said it used AlphaProteo together with a SynthID-enabled version of ProteinMPNN to create watermarked protein binders, which are proteins designed to attach to specific biological targets. For structure prediction, DeepMind said it fine-tuned part of AlphaFold 3’s diffusion network so predicted 3D protein coordinates carry a detectable signature. The company reported wet-lab testing on three targets: VEGF-A, SC2RBD — the receptor-binding domain of the SARS-CoV-2 spike protein — and PD-L1.

DeepMind said the watermarked designs matched unwatermarked versions on hit rate, binding affinity and natural sequence diversity in those tests. It also said SynthID Bio offers “near-perfect detectability” for AlphaFold 3 predictions while preserving prediction accuracy and staying robust to minor coordinate noise. The practical use case, as DeepMind framed it, is provenance: giving model developers, DNA synthesis providers and database curators a way to tell when a sequence or predicted structure came from an AI system. In the blog post, the company said the approach could help DNA synthesis firms screen orders, help curators flag AI-designed entries in public biological databases and strengthen layered biosecurity defenses. Sarah Carter, quoted in the post, called SynthID Bio “an important piece of the puzzle for tracking the provenance of biological designs.”

DeepMind said Adaptyv Bio helped with in vitro validation. The company also said it worked with the Hie lab at Stanford University and the Arc Institute on applying watermarking to bacteriophage DNA, with early lab testing in bacterial cultures indicating the watermarked bacteriophages remained functional. DeepMind said it is publishing a methods paper, open-sourcing code and in vitro data, and releasing weights to the research community. But at the time of an initial check summarized in the research report behind this assignment, no linked public methods manuscript or dedicated public code repository was identified from the announcement page itself, so those materials should be understood as the company’s stated plan rather than confirmed public releases.

The announcement extends DeepMind’s SynthID branding beyond media generation. The company has previously used SynthID for watermarking images, audio and text, and this is its push into synthetic biology. But the broader idea of protein watermarking is not entirely new. Academic researchers published FoldMark in 2025, describing protein watermarking with wet-lab validation in other protein systems. That means DeepMind should not be credited with creating the first watermarked functional proteins outright. Its narrower claim, as presented in the blog post, is that these are the first watermarked and biologically functional protein binders.

DeepMind also cast the work as part of a larger safety push around AI-generated biology. In the company’s blog post, Carter said, “By linking designs to the model developer, these watermarks empower developers to lead on safety and allow synthesis providers to streamline screening for customers who have used those models.” The post also quoted James Diggans, saying, “For Twist, watermarking offers a promising new addition to the biosecurity toolbox that could strengthen screening, focus resources on sequences that warrant closer review and make biosecurity more efficient as AI-designed biology continues to advance.” DeepMind said a major open challenge is making the watermark more robust against deliberate tampering, and it said broader benefits will require collaboration across the research and biosecurity communities.