Google DeepMind open‑sources SynthID Bio to watermark AI‑designed proteins
DeepMind says its Nature paper shows watermarks for protein sequences and AlphaFold 3 structures, and the detection code and validation data are now on GitHub.
Google DeepMind published SynthID Bio in Nature on Wednesday, a set of methods for embedding watermarks in AI-designed proteins, and released code for it. Chief executive Demis Hassabis said on X that the aim is biosecurity, so that AI-generated proteins can be watermarked, and that the tools are open source so the research community can build on the work.
https://x.com/demishassabis/status/2105348732464070823
The GitHub repository describes two parts. SynthID Bio-Sequence embeds a watermark while a protein is being designed, and its README calls it a drop-in replacement for existing design pipelines, built on ProteinMPNN. SynthID Bio-Structure is a modified AlphaFold 3 variant that adds watermarks to predicted 3D structures. A standalone calculator scores how detectable a watermark is in a FASTA file.
The release includes the watermarked ProteinMPNN code, detection tools, a test suite and the validation data behind the paper. DeepMind's own code is under Apache 2.0, ProteinMPNN keeps its MIT licence, and AlphaFold 3 weights need separate terms. The ProteinMPNN training code and weights are not included. The repository had 2 stars when the paper read it.
Published summaries of the paper say a watermark in an AI-designed protein sequence survived physical synthesis and the protein still bound its target, which DeepMind is said to call the first watermarked, biologically functional binders. The paper itself sits behind a login, so the paper has not read the results and has not seen them checked by outside labs.
The summaries say the intended use is gene synthesis screening, where a watermark could mark a sequence as coming from a model with safeguards and let it be fast-tracked. Whether screening providers or regulators will adopt the scheme is not yet known, and a watermark may not help against proteins designed with tools that do not use it.