How SynthID Bio Watermarks AI-Designed Proteins—and What Its Tests Show

SynthID Bio is Google DeepMind’s proof of concept for making an AI-designed protein carry a detectable signal, rather than relying only on a separate label that could become detached from the design. DeepMind says its laboratory tests preserved biological function, while noting that resisting deliberate watermark removal remains an open challenge. 6

How SynthID Bio embeds and detects a watermark

The method differs by what is being marked. For protein sequences, the watermark is incorporated into the amino-acid choices made during design. A detector then checks the sequence for the signal. For predicted 3D structures, DeepMind fine-tunes the structure-generating model so its predicted atomic coordinates carry a detectable signature. 6

A sequence-based watermark can remain in a synthesized protein because the protein is built from that sequence. A signal applied only to a predicted 3D structure, however, should not be assumed to carry over into a physical protein. 6

What the reported tests found

DeepMind reports that synthesized, watermarked protein binders targeting VEGF-A, the SARS-CoV-2 spike receptor-binding domain, and PD-L1 remained functional and performed comparably to unwatermarked designs. It also says watermarking AlphaFold 3 predictions preserved model performance. 6

In separate early tests, watermarked bacteriophages designed with Evo 2 were functional in bacterial culture. DeepMind says fuller technical details of that work are still forthcoming, so the finding should be read as an early result rather than a complete evaluation. 6

How the watermark could support screening and provenance

DNA-synthesis providers could use a detectable design-origin signal as an additional reason to flag an order for review, alongside existing screening. The watermark is a possible extra signal, not a replacement for screening practices. 6

Public biological databases—including the Protein Data Bank, UniProt, and GenBank—could also check submissions for a watermark. That could help label AI-generated entries or send them for review, supporting clearer provenance. These are potential uses, not evidence that the watermark alone can establish a design’s full history. 6

What DeepMind is making available—and what remains uncertain

DeepMind says it is publishing its methods paper, open-sourcing code and in vitro data, and releasing model weights to researchers. 6

The central limitation is that a detectable watermark is not automatically an indelible one. SynthID Bio is a proof of concept: DeepMind reports promising function-preservation results, but deliberate tampering remains an open challenge. 6