Google DeepMind Introduces SynthIDBio for Watermarking AI-Designed Proteins

Google DeepMind has unveiled SynthIDBio, a novel method for embedding invisible signatures directly into AI-designed biological sequences and their three-dimensional structures. This development, detailed in an October 1, 2026 report, aims to address emerging biosecurity risks and the potential spread of misleading scientific data associated with AI-generated molecules.

SynthIDBio operates by integrating digital patterns, either as a secret code within the amino acid order for protein sequences or as an invisible pattern in the 3D coordinates of atoms for biomolecular structures. Experiments demonstrated that watermarked protein binders, targeting molecules such as the coronavirus spike protein and human immune regulators, attached with comparable strength to their non-watermarked counterparts. A detection tool successfully identified all watermarked sequences and over 99.8% of watermarked 3D structures, without significantly affecting their function or accuracy.

The technology, whose details were published in the journal Nature by David Stutz et al. in 2026, offers a proof of concept for function-preserving biological watermarking. While the 3D structure watermark can be susceptible to removal during standard structural tweaks like relaxation, researchers anticipate this limitation can be addressed through further training. The open-source code for SynthIDBio-sequence is available on GitHub.

What to watch: Further developments in SynthIDBio's robustness against structural modifications.

Editor's note: Comprehensive summary that correctly incorporates technical details, study findings, and limitations from the full text.

This article is AI-generated and fact-gated. Original reporting: Phys.org