DeepMind publishes SynthID Bio in Nature: first verifiable watermark for AI-designed proteins

Google DeepMind published SynthID Bio in Nature on September 30, extending the SynthID watermarking family first introduced for images, video, audio, and text in 2023 into synthetic biology. SynthID Bio is the first verifiable watermarking method for AI-designed proteins and comes in two complementary parts: SynthID Bio-sequence subtly steers amino acid selection during sequence generation in ProteinMPNN under a secret key, leaving a statistical pattern that a detector can read; SynthID Bio-structure fine-tunes a small part of the AlphaFold 3 diffusion network so the watermark lives in the model weights, automatically marking every AlphaFold 3 output.

DeepMind also published wet-lab validation. The team combined AlphaProteo with a SynthID Bio-enabled ProteinMPNN to design protein binders against three targets: VEGF-A, the SARS-CoV-2 spike receptor-binding domain, and PD-L1. Watermarked designs matched unmarked ones on hit rate, binding affinity, and sequence diversity, producing low-nanomolar binders against the SARS-CoV-2 RBD and sub-nanomolar binders against VEGF-A and PD-L1. On AlphaFold 3, watermarking had minimal impact on structure prediction accuracy and detection rates approached perfect, surviving simple translations, rotations, and small coordinate noise.

The primary use case is DNA-synthesis screening. Synthesis providers normally compare customer sequences against databases of known threats, but AI can now generate sequences with little resemblance to known hazards while retaining dangerous function, eroding the screening layer. SynthID Bio provides provenance: if a protein comes from a watermarked, safety-controlled model, synthesis providers can read the watermark in the encoded sequence and confirm its origin. DeepMind positions it as one layer in a Swiss-cheese defense alongside model-side risk mitigation, customer vetting, and sequence-database screening.

DeepMind is publishing the methods paper, releasing the code and lab data, and sharing model weights with the research community. The team has collaborated with the Hie lab at Stanford and the Arc Institute to watermark genomes as well, embedding a watermark in a bacteriophage genome designed by Evo 2, with early tests confirming the marked phage remains functional.