AI-designed proteins can carry a watermark and still bind their targets

Google DeepMind's SynthID Bio announcement describes a way to make part of that provenance travel within the design itself: a detectable pattern in the amino-acid sequence. The September 30 research gives laboratories a concrete starting point for identifying designs from participating models without relying only on a removable file label.

The peer-reviewed Nature study tested binders against three protein targets. It began with 15 previously successful structural backbones per target, then generated new sequences with or without a watermark. Across 756 sequences, excluding parent designs and controls, the researchers found no significant population-level shift in binding affinity. This is evidence from selected, already promising structures; it does not establish that watermarking preserves every protein's function.

One qualification is visible in the results. At the looser binding threshold shown below, 150 of 222 unwatermarked designs were hits, compared with 155 of 267 designs using one watermark setting. That difference was statistically significant. At the stricter threshold, the study found no significant hit-rate differences. The other practical limit is deliberate removal: the researchers showed that resequencing can erase the signal.

The useful advance is an experimentally tested provenance mechanism. A positive watermark could help synthesis providers or database curators identify a design's origin. It cannot certify that the protein is safe, and an absent watermark cannot establish a natural origin. DeepMind says wider use still requires further research and cooperation across the field.