DeepMind's SynthID Bio watermarks AI-designed proteins and DNA

Google DeepMind has extended its SynthID watermarking system from text and images into biology. The new version, SynthID Bio, hides a detectable signature inside AI-generated biological sequences without changing what they do. The problem it targets is specific: modern design models can produce proteins that work but look nothing like anything in existing threat databases, so the screening tools that DNA synthesis providers rely on struggle to flag them. Public sequence databases also risk filling up with AI-generated entries that nobody labeled.

The method adapts to the kind of data. For protein sequences it nudges the choice of amino acids; for predicted 3D structures it shifts atomic coordinates slightly; for genomic DNA it fine-tunes the diffusion weights inside AlphaFold 3 so the signature comes out in the model's own generations. DeepMind reports that watermarked protein binders kept the same hit rate, binding affinity, and sequence diversity as unmarked ones across three targets: VEGF-A, the SARS-CoV-2 spike RBD, and PD-L1. The team describes these as the first watermarked protein binders that are still biologically functional, and says watermarked bacteriophages designed by the Evo 2 model also worked in early lab tests. You can read the write-up for the per-target results.

Why it matters

If you run a DNA synthesis service or maintain a sequence database, this gives you a way to tell AI-generated material apart from natural sequences without slowing down legitimate research, as long as the design tools you screen against actually embed it.