source: Google DeepMind: Introducing SynthID Bio
level: research
google deepmind has introduced synthid bio, a watermarking system for synthetic biology. it embeds an imperceptible signature directly into the biological code of ai-designed proteins. the watermark is verifiable on both the digital model and the physical protein. this approach aims to strengthen biosecurity and scientific integrity. it adapts to different data types, guiding amino acid choices for sequences and adjusting atomic coordinates for 3d structures. the goal is to track provenance without harming function.
in wet-lab tests, watermarked protein binders matched the hit rate, binding affinity, and natural sequence diversity of unwatermarked versions. tests covered three targets: vegf-a, sars-cov-2 spike protein rbd, and pd-l1. for protein folding, synthid bio fine-tunes part of alphafold 3's diffusion network. this builds watermarking into model weights, so predicted 3d coordinates carry a detectable signature. the method preserves alphafold 3 accuracy and offers near-perfect detectability, even under digital noise or minor coordinate changes.
the watermark could help dna synthesis providers screen orders more efficiently. ai can create sequences unlike known hazards, making manual review slow. synthid bio provides an automated verification signal that an order came from a trusted model. it may also help label synthetic entries in public databases like the protein data bank. google deepmind is open-sourcing the code, in vitro data, and model weights. ongoing work with stanford and arc institute extends watermarking to bacteriophage genomes.
why it matters: watermarking ai-generated proteins helps track their origin, making biosecurity screening faster and reducing database contamination.