Google DeepMind tests watermarks for AI-designed proteins
SynthID Bio marks generated protein sequences and predicted structures; DeepMind reports functional binders in early lab tests.
What DeepMind showed
Google DeepMind introduced SynthID Bio on September 30 as a family of methods for watermarking AI-generated biological designs. One method steers the amino-acid choices in a generated protein sequence so a detectable signature remains in the synthesized molecule. Another modifies a protein-structure prediction model so its three-dimensional coordinates carry a signature. These are different forms of evidence: a marked sequence could be checked in a physical protein, while a marked structure helps identify an AI-generated prediction.
DeepMind says it tested watermarked protein binders against three targets in the laboratory and found that the designs retained measures of binding and diversity comparable with unwatermarked versions. It also reports that its modified AlphaFold 3 predictions retained accuracy while carrying a detectable signal. These are the company’s experimental results; the reviewed source does not establish independent replication across all protein types or design methods.
Why provenance is only one layer
The company proposes that a watermark could help DNA synthesis providers check where an unfamiliar order originated and help scientific databases label synthetic entries. That would be a provenance signal, not proof that an order or protein is harmless. DeepMind explicitly identifies deliberate tampering as an unresolved challenge and describes the technology as one layer among other biosecurity controls. The work is a research proof of concept, not a claim that universal screening is already deployed.