Google DeepMind's SynthID Bio Watermarks AI-Designed Proteins

SynthID Bio, released by Google DeepMind on September 30, 2026 (local time), is a biology-specific watermark that embeds an identifiable signal into AI-designed proteins without compromising their function, with the signal remaining detectable even in synthesized molecules. In experiments where more than 1,300 binders were actually synthesized, watermarked binders showed no significant difference in binding success rate or binding affinity compared to non-watermarked proteins.

How does SynthID Bio work?

It is designed to make subtle adjustments to a protein's amino acid sequence or three-dimensional atomic coordinates, embedding an invisible watermark while preserving the protein's original biological function. The key point is that the signal can be detected not only in digital models but also in physically synthesized protein molecules.

• Sequence watermarkingThis involves making minor substitutions among amino acids with similar properties so as not to harm biological function. It was tested by applying it to the protein design model AlphaProteo and the protein sequence generation model ProteinMPNN.
• Synthesis experimentsMore than 1,300 binders were synthesized targeting three proteins: vascular endothelial growth factor (VEGF-A), the receptor-binding domain (RBD) of the COVID-19 virus spike protein, and PD-L1. No significant differences appeared in binding success rate, binding affinity, or natural sequence diversity, and the watermark signal was identified in physical molecular samples using actual protein analysis equipment such as mass spectrometers.
• Structural watermarkingPart of AlphaFold 3's diffusion network was fine-tuned to embed watermarking capability directly into the model weights. Google DeepMind stated that the loss in structure prediction accuracy was kept under 1%, and the watermark was verified with over 99% accuracy even after sequence modifications or noise were applied.
• Genome applicationA watermark was embedded in a bacteriophage genome designed by the genome design model Evo 2, and bacterial culture experiments were conducted. Early results reportedly confirmed that function was preserved.

What changes for biosecurity?

Google DeepMind explained that SynthID Bio could also be used for biosecurity purposes. Existing security screening by DNA synthesis companies relies heavily on comparing sequences against known hazardous materials, but when AI generates novel proteins or genomes that are not similar to existing biological sequences, this approach alone makes it difficult to determine origin and risk. Google DeepMind presented two use cases.

• Serving as a signal during DNA synthesis to confirm that a design was generated by an AI model, helping identify orders that require additional review
• Marking AI-generated data as synthetic or classifying it for separate review when registered in biological databases such as the Protein Data Bank (PDB), UniProt, and GenBank

Google DeepMind also acknowledged limitations, emphasizing that SynthID Bio is not a single solution that resolves all biosecurity concerns. Making the technology more robust against malicious actors who might deliberately tamper with or remove the watermark remains a challenge. The company is also researching ways to combine the watermark with metadata systems that link it to provenance information, as well as central repositories that manage AI-generated biological data. It is collaborating on biosecurity research with teams from Stanford University, the Arc Institute, and UC Berkeley, and has open-sourced its research paper along with code and experimental data, while releasing research model weights on GitHub.

Implications for AX in Korean Enterprises

What stands out in this release is the design choice to embed provenance labeling for AI-generated content at the generation stage itself, rather than relying solely on post-hoc inspection of the output. Screening methods that compare against known sequences lose their reference point when confronted with sequences an AI has generated for the first time. SynthID Bio embeds watermarking capability into AlphaFold 3's weights so that the signal is carried in the model's output from the very start.

Domestic bio and pharmaceutical organizations using AI design models for research can directly verify, using the released code and research weights, how the watermark affects protein function under their own experimental conditions. Even for organizations outside the bio field, the question is the same: when AI-generated designs or data move into internal databases or external partner systems, it is worth first checking at what stage a label identifying it as AI-generated content is actually applied.

Source: Google DeepMind unveils watermarking technology to track AI-generated proteins and genomes