Introducing SynthID Bio
Proof of concept for watermarking AI-generated proteins while preserving biological function.
Pushmeet Kohli, David Stutz, Ali Cowen-Rivers and Jeremy Ratcliff have unveiled SynthID Bio, a watermarking technology designed to address challenges posed by generative AI in synthetic biology. The system integrates an imperceptible signature into biological code, ensuring that watermarks can be verified both digitally and in physical proteins without compromising their functionality.
SynthID Bio tackles the issue of AI-generated proteins potentially bypassing traditional DNA synthesis screening and misleading public databases. By subtly adjusting amino acids for sequences and atomic coordinates for 3D structures, the technology preserves biological function during lab testing. Experiments demonstrated that watermarked protein binders matched the hit rate, binding affinity, and natural sequence diversity of unwatermarked versions.
For protein folding, SynthID Bio fine-tunes AlphaFold 3's diffusion network, embedding watermarking directly into the model's weights. This preserves prediction accuracy while ensuring detectability. The watermarking approach acts as a verification layer within the biological design, linking designs to model developers and enabling efficient DNA synthesis screening.
Major stakeholders, including biosecurity experts and DNA synthesis providers, recognize the potential of SynthID Bio in enhancing biosecurity and scientific integrity. However, challenges remain in making the watermark robust against deliberate tampering and integrating it with provenance metadata approaches for better identification and tracking of AI-generated proteins.
Written by urgent.news from Google DeepMind's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
