Sakhanda Wire
NVDA $154.51 -1.04% MSFT $506.49 -2.22% GOOGL $281.27 -0.86% META $621.51 -0.78% AMZN $256.90 -0.33%
← Back to the news

SynthID Bio introduces watermarking for AI-designed proteins

A new proof of concept called SynthID Bio embeds an imperceptible, verifiable signature into AI-generated biological designs. The watermark is intended to remain detectable not only in a digital model, but also in a synthesized protein, while preserving biological function in laboratory tests.

The system adapts to different data types by subtly influencing amino-acid choices in sequences or modifying atomic coordinates in predicted three-dimensional structures. In wet-lab tests involving protein binders for VEGF-A, the SARS-CoV-2 spike protein RBD and PD-L1, watermarked designs matched unwatermarked versions in hit rate, binding affinity and natural sequence diversity. A version integrated into AlphaFold 3 also preserved prediction accuracy and showed near-perfect detectability under digital noise and minor coordinate changes.

The developers say the technology could add a verification layer to DNA-synthesis screening and help identify synthetic entries in databases including the Protein Data Bank, UniProt and GenBank. They caution that the watermark is not a standalone biosecurity solution and are studying resistance to deliberate tampering. The project’s methods paper, code and in vitro data are being released, while weights are being made available to researchers. Early tests also found that bacteriophages designed with Evo 2 and watermarked through related work remained functional in bacterial cultures.

Written by Sakhanda Wire AI from reporting by DeepMind. Facts, quotes and full context are in the original article below.

Originally published by DeepMind on

Read the original on DeepMind ↗

The reporting and the image belong to DeepMind and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news