SynthID Bio Watermarks AI-Designed Proteins for Biosecurity

Google DeepMind has introduced SynthID Bio, a way to hide an invisible watermark inside AI-designed proteins and DNA so that the origin of a design can be checked later. Announced on September 30, 2026, it extends the SynthID idea from images and audio into biology, and DeepMind calls the current work a proof of concept that is backed by a methods paper in Nature.

• SynthID Bio embeds an imperceptible signal in protein sequences by subtly biasing which amino acids a model picks, and it also adjusts predicted 3D coordinates.
• It works with AlphaFold 3 and ProteinMPNN, and was extended with Stanford's Hie Lab and the Arc Institute to Evo 2 for bacteriophage DNA.
• In wet-lab tests on three targets, watermarked binders matched unwatermarked ones on hit rate, affinity and diversity, according to Google.
• Code, in vitro data and model weights are being released to the research community.

What is SynthID Bio?

SynthID Bio is a watermarking technique for generative biology. When a protein design model chooses each amino acid, the method nudges those choices in a statistical pattern that a detector can recognize later but that does not change how the protein behaves. For structure prediction, it adjusts the 3D coordinates in a similarly quiet way. The result is a fingerprint that travels with the design.

It builds on the approach we covered when SynthID came to GPT-Live audio, applied this time to molecules instead of media.

Does watermarking hurt the science?

This is the question that decides whether researchers will adopt it. Google says it tested watermarked binders against three targets, VEGF-A, the SARS-CoV-2 spike receptor-binding domain and PD-L1, with lab validation from Adaptyv Bio. According to Google, the watermarked designs matched unwatermarked ones in hit rate, binding affinity and diversity. With Stanford's Hie Lab and the Arc Institute, the team also extended the method to Evo 2 and watermarked bacteriophage DNA, with early bacterial-culture tests showing the genomes stay functional. These are Google's own results, so independent replication will be the next milestone.

How could it help DNA synthesis screening?

The practical goal is to help DNA synthesis providers check whether an order came from a trusted model. Google's announcement names Twist Bioscience as a collaborator. A provider that can read the watermark gets an extra signal when deciding how to handle a request, and researchers get a way to show which model produced their design.

DeepMind is releasing code, data and model weights for the community, which lowers the barrier for other labs to try the method and report back. The company is open that this is an early, proof-of-concept stage, and it positions the work as one layer among several rather than a complete answer.

Why this matters

Generative biology is moving quickly, and open tools are a big reason. Provenance features like SynthID Bio aim to let that openness continue while giving providers and researchers better ways to trace designs. For a broader view of where AI is being applied to science, see our AI coverage.

Sources: Google: Introducing SynthID Bio — September 30, 2026; Nature: SynthID Bio methods paper — September 30, 2026; Science: Method to watermark AI-designed proteins — September 30, 2026.