Google DeepMind has introduced SynthID Bio, a family of watermarking methods for AI-generated proteins and their predicted 3D structures, and is releasing the code, lab data and model weights to the research community. The methods paper was published in Nature on September 30.

The idea is to embed a detectable signature in the biological code itself — subtly steering amino-acid choices for sequences, adjusting atomic coordinates through a fine-tuned slice of AlphaFold 3 — so the mark can be verified on the physical protein after wet-lab manufacturing, not just in a digital file. DeepMind says watermarked protein binders matched unwatermarked ones on hit rate, binding affinity and sequence diversity in lab tests.

The biosecurity case is the pitch: DNA synthesis companies screen orders for threats, and AI can now generate entirely novel sequences that slip past conventional filters. A watermark would give screeners an automated signal that an order came from a trusted model. DeepMind concedes it is not tamper-proof — a resequencing attack removed the sequence mark — and frames it as one layer alongside model mitigations and customer vetting.