DeepMind’s protein watermark aims to make AI-designed molecules easier to identify

Google DeepMind’s SynthID Bio embeds a detectable pattern in AI-designed proteins while aiming to preserve their function. In reported lab tests, watermarked proteins bound to three targets as strongly as unmarked ones, a promising proof of concept for identifying designs from known AI systems.

Google DeepMind Watch analysis

What happened

The method nudges amino-acid choices as a protein is designed, rather than adding a mark afterwards. According to The Angry Optimist’s account of a study in Nature, the detector found the watermark in more than 99.8% of marked structures while wrongly flagging no more than one in a thousand unmarked structures.

Researchers tested proteins designed to bind to three targets, including part of the COVID-19 virus’s spike protein. They reported no statistically significant difference in binding strength between marked and unmarked proteins. The report also describes early tests of watermarked AI-designed bacteriophages in bacterial cultures, where the marked phages still worked.

Why it matters

DNA manufacturers screen sequences for known threats, but AI-designed sequences may not resemble anything in those databases. A reliable mark from a known AI system could help distinguish its designs from sequences of unknown origin, potentially letting screening teams focus more attention where provenance is unclear.

That is a useful idea, not a replacement for screening. The reported tests cover a limited set of proteins and early phage work; they do not establish how the method performs across all designs, models or real-world ordering systems.

Our read

The clever part is that the watermark is designed into the protein rather than pasted on afterwards. If it holds up beyond these proof-of-concept tests, it could give biosecurity screening a practical provenance signal without making the marked proteins less functional in the tests described. The Angry Optimist reports that DeepMind is sharing code and lab data, while withholding the detector and watermark keys.

What to watch

• Whether independent groups reproduce the detection and binding results across more proteins.
• How the watermark performs when designs come from different models or are altered after creation.
• Whether DNA manufacturers can incorporate provenance signals into screening without treating an absent mark as proof of safety.

Discussion spark: Should DNA manufacturers prioritise faster approval for designs carrying a recognised AI watermark, or should every sequence face the same level of screening?

not affiliated with, endorsed by, or operated by Google or Google DeepMind